Skip to main content
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 Nov 1.
Published in final edited form as: Med Image Anal. 2026 Jul 31;114:104234. doi: 10.1016/j.media.2026.104234

KPIs 2024 challenge: Advancing glomerular segmentation from patch- to slide-level

Ruining Deng a,b, Tianyuan Yao a, Yucheng Tang c, Junlin Guo a, Siqi Lu a, Juming Xiong a, Lining Yu a, Quan Huu Cap e, Pengzhou Cai f,g,h, Libin Lan f, Ze Zhao i, Adrian Galdran j, Amit Kumar k, Gunjan Deotale k, Dev Kumar Das k, Inyoung Paik l, Joonho Lee l, Geongyu Lee l, Yujia Chen m, Wangkai Li m, Zhaoyang Li m, Xuege Hou n, Zeyuan Wu n, Shengjin Wang n, Maximilian Fischer o, Lars Krämer o, Anghong Du p, Le Zhang p, Maria Sánchez Sánchez j, Helena Sánchez Ulloa j, David Ribalta Heredia j, Carlos Pérez de Arenaza García j, Shuoyu Xu q, Bingdou He q, Xinping Cheng r, Tao Wang r, Noémie Moreau s, Katarzyna Bozek s, Shubham Innani t, Ujjwal Baid t, Kaura Solomon Kefas u, Bennett A Landman a, Yu Wang d, Shilin Zhao d, Mengmeng Yin d, Haichun Yang d, Yuankai Huo a,*
PMCID: PMC13623399  NIHMSID: NIHMS2210438  PMID: 42612515

Abstract

Chronic kidney disease (CKD) is a major global health issue, affecting over 10% of the population and causing significant mortality. While kidney biopsy remains the gold standard for CKD diagnosis and treatment, the lack of comprehensive benchmarks for kidney pathology segmentation hinders progress in the field. To address this, we organized the Kidney Pathology Image Segmentation (KPIs) Challenge, introducing a dataset that incorporates preclinical rodent models of CKD with over 10,000 annotated glomeruli from 50 Periodic Acid Schiff (PAS)-stained whole slide images. As a preclinical benchmark, this challenge is designed to evaluate and compare computational methods under controlled experimental conditions. The challenge includes two tasks, patch-level segmentation and whole slide image segmentation and detection, evaluated using the Dice Similarity Coefficient (DSC) and F1-score. By encouraging innovative segmentation methods that adapt to diverse CKD models and tissue conditions, the KPIs Challenge aims to advance kidney pathology analysis, establish new benchmarks, and enable precise, large-scale quantification to support preclinical disease research and future translation to human kidney pathology.

Keywords: Segmentation, Renal pathology, MICCAI challenge

1. Introduction

Chronic kidney disease (CKD) represents a significant global health challenge, causing more deaths annually than breast and prostate cancer combined (Cirillo et al., 2024; Iglesias et al., 2012). Affecting over 10% of the global population, CKD impacts approximately 800 million individuals worldwide. Kidney biopsy, through both open and percutaneous methods, remains the gold standard for diagnosing and guiding the treatment of CKD (Hogan et al., 2016; Agarwal et al., 2013). In pathological image analysis, especially within the context of kidney disease, tissue segmentation plays a pivotal role in enabling large-scale, computer-assisted quantification in pathology (Bengtsson et al., 2017; Marti-Aguado et al., 2021; Gomes et al., 2021). This analysis provides clinical value for disease diagnosis (Mounier-Vehier et al., 2002), severity assessment (Kellum, 2008), and treatment effectiveness evaluation (Jiménez-Heffernan et al., 2006).

The advent of deep learning has revolutionized kidney pathology image segmentation (Kumar et al., 2017; Ding et al., 2020; Ren et al., 2017; Bel et al., 2018; Zeng et al., 2020), yet significant challenges remain. Chief among these is the lack of comprehensive benchmarks for developing and evaluating segmentation methods. Existing public datasets for kidney pathology segmentation are limited, primarily containing samples from healthy patients.1,2 This limitation arises because tissue samples are typically obtained through needle biopsies, yielding only small samples. As a result, most publicly available datasets do not capture the full spectrum of pathological variability observed in CKD, limiting their utility for benchmarking robust segmentation methods. Consequently, there is a pressing need to establish extensive kidney pathology datasets encompassing diverse CKD models.

To address these limitations, we organized the Kidney Pathology Image Segmentation (KPIs) Challenge, expanding the dataset by incorporating preclinical animal models, specifically whole kidney sections from diseased rodents. This challenge is designed as a preclinical benchmark to enable controlled, large-scale evaluation of segmentation algorithms. This competition aims to identify state-of-the-art segmentation methods for glomerular identification across various CKD models. Participants are tasked with developing algorithms capable of accurately segmenting glomeruli at the pixel level, adapting to different CKD models and tissue conditions. This includes distinguishing glomeruli from surrounding tissue components under diverse preparation scenarios, showcasing both the versatility and precision of their approaches. Furthermore, the challenge encourages innovative solutions to address potential obstacles such as variations in glomeruli size, shape, and structural integrity resulting from disease states or preparation techniques. Insights gained from this preclinical setting are intended to inform future studies on human kidney pathology, including recent efforts on cross-species analysis between mouse and human kidney samples, and to facilitate subsequent translational validation (Yu et al., 2024; Zhu et al., 2025).

To comprehensively evaluate participants’ models, the challenge features two distinct tasks (as shown in Fig. 1) aimed at advancing the field of kidney pathology image analysis:

Fig. 1.

Fig. 1.

Overview of the challenge task setup. The KPIs Challenge involves two tasks: (1) Patch-level Diseased Glomeruli Segmentation, focusing on the precise delineation of glomeruli in high-resolution image patches, and (2) WSI-level Diseased Glomeruli Segmentation, requiring robust segmentation across entire tissue sections. Example images illustrate the input data, ground truth masks, and overlay visualizations for each task. The dataset includes diverse mouse models to highlight morphological and pathological variations.

  1. Patch-Level Segmentation: Segmentation of glomeruli within specific image patches.

  2. WSI-Level Segmentation and Detection: Segmentation and detection of glomeruli across entire kidney slide images.

Model performance is assessed using the Dice Similarity Coefficient (DSC) and F1-score on the testing dataset. While patch-level evaluation focuses on local segmentation accuracy, the inclusion of WSI-level analysis is intended to provide a more challenging and discriminative assessment of model robustness and scalability under realistic conditions.

The KPIs Challenge encompasses a broad spectrum of kidney disease models, including normal and multiple CKD-specific conditions, derived from preclinical rodent models. As a pioneering effort within the MICCAI community, the challenge features an extensive collection of 10,000 normal and diseased glomeruli from 50 Periodic Acid Schiff (PAS)-stained whole slide images (WSIs). Each image includes nephrons, with each nephron containing a glomerulus and a small cluster of blood vessels. Participants are tasked with developing algorithms that can accurately segment glomeruli at a pixel level. To the best of our knowledge, this is the first MICCAI challenge exclusively dedicated to segmenting functional units in kidney pathology across diverse CKD models.

2. Related works

2.1. Patch-level glomerular segmentation

Patch-level glomerular segmentation has been a focal point in renal pathology, aiming to delineate glomeruli within small, manageable regions of kidney histology images (Ginley et al., 2019). Early methodologies predominantly utilized classical image processing techniques, including thresholding, edge detection, and morphological operations (Altini et al., 2020; Ginley et al., 2017), to identify glomerular structures. While these approaches established foundational workflows, they often encountered challenges due to the intricate morphology of glomeruli and variability in staining methods, leading to inconsistent segmentation outcomes (Altini et al., 2020).

The advent of deep learning, particularly convolutional neural networks (CNNs), has significantly advanced the field of medical image segmentation (Gallego et al., 2021). Architectures such as U-Net and its derivatives have been extensively applied to patch-level glomerular segmentation, demonstrating enhanced accuracy and efficiency by learning complex patterns inherent in kidney tissues (Singh Samant et al., 2023; Gallego et al., 2021; Deng et al., 2023b). Despite these advancements, several challenges persist in patch-level segmentation. A primary concern is the lack of global contextual information, as models trained on isolated patches may not effectively capture the spatial relationships and broader tissue architecture present in WSIs (Wu and Moeckel, 2023). This limitation can lead to inaccuracies, particularly when glomeruli are located near the periphery of patches or exhibit atypical presentations. Additionally, variability in image quality, staining techniques, and the inherent complexity of glomerular structures can adversely affect segmentation performance (Wu and Moeckel, 2023). To address these issues, recent research has explored hybrid models that integrate CNNs with transformer-based architectures to enhance both local feature extraction and global context understanding (Liu, 2024; Yin et al., 2024; Wu and Moeckel, 2023; Deng et al., 2024b,a, 2023a; Cui et al., 2024). For example, a hybrid CNN-TransXNet (Liu, 2024) approach has been proposed to improve segmentation accuracy by combining the strengths of CNNs in capturing fine-grained local features with the global contextual understanding provided by transformers.

2.2. Whole slides image-level glomerular detection and segmentation

WSI glomerular detection and segmentation have become pivotal in computational kidney pathology, enabling comprehensive analysis of kidney tissues at high resolution. Unlike patch-based methods, WSI approaches consider the entire tissue context, facilitating more accurate localization and characterization of glomeruli (Bueno et al., 2020; Tang et al., 2024). Beyond glomerular-focused studies, digital pathology approaches have also been applied to kidney-wide histopathology analysis using ensemble deep learning frameworks, such as the Dartmouth Kidney Cancer Histology Dataset, which highlights the effectiveness of ensemble models in complex renal tissue classification tasks (Jain et al., 2025).

Recent advancements have introduced holistic frameworks that integrate detection, segmentation, and lesion characterization within unified pipelines. For instance, the Glo-In-One (Yao et al., 2022; Yu et al., 2024) toolkit performs comprehensive glomerular quantification from WSIs, streamlining the analysis process for non-technical users. Extension plugin GloFinder (Yue et al., 2024b), designed for QuPath, enables single-click automated glomeruli detection across entire WSIs, streamlining the analysis process for non-technical users.

Additionally, deep learning models have been employed to enhance glomerular detection in WSIs, such as modified U-Net architecture (Jha et al., 2021). These models leverage convolutional neural networks to accurately identify glomeruli, demonstrating significant improvements over traditional methods (Jha et al., 2021; Jiang et al., 2021).

Despite these advancements, several challenges persist in WSI glomerular analysis. Processing gigapixel WSIs demands substantial computational resources, necessitating efficient algorithms and hardware (Becker et al., 2020). Differences in staining protocols, imaging modalities, and tissue preparation can introduce variability, affecting model performance (Ehteshami Bejnordi et al., 2016). Obtaining pixel-level annotations for WSIs is labor-intensive, often resulting in limited training data for supervised learning models (Brixtel et al., 2022).

To address challenges in WSI glomerular detection and segmentation, researchers have developed innovative strategies. For example, Weighted Circle Fusion (WCF) (Yue et al., 2024a) has been proposed to enhance detection precision by fusing outputs from multiple models using confidence-weighted circle representations. This method not only improves accuracy but also reduces false positives by effectively merging overlapping detections. Furthermore, the human-in-the-loop (HITL) approach has significantly improved annotation efficiency, combining machine learning detection with human verification. However, WCF is not without limitations. It demonstrates slightly lower recall compared to methods like Non-Maximum Suppression (NMS), and its reliance on careful parameter tuning may restrict generalizability across different datasets (Yue et al., 2024a).

2.3. Pathology challenges and benchmark datasets

Community-driven challenges have played an important role in advancing medical image analysis by providing standardized datasets, evaluation protocols, and objective comparisons of competing methods. Notable examples include challenges in digital pathology focusing on tissue segmentation, cell detection, and disease classification across multiple organs and modalities (Bandi et al., 2018; Bulten et al., 2022a; Kumar et al., 2019). For instance, CAMELYON16/17 benchmarks whole-slide lymph node metastasis detection, enabling standardized comparison of WSI-level tumor detection methods (Litjens et al., 2018; Ehteshami Bejnordi et al., 2017; Bándi et al., 2019). Slide-level grading challenges such as PANDA (Bulten et al., 2022b), evaluate prostate cancer detection and Gleason grading at scale, accelerating progress in robust slide-level classification and grading. Moreover, the MoNuSeg 2018 challenge (Kumar et al., 2020) focuses on multi-organ nuclei instance segmentation, supporting objective evaluation of cell-level segmentation methods across diverse tissue types. Collectively, these efforts have significantly contributed to methodological progress and reproducibility in the field.

Within renal pathology, however, existing benchmarks are relatively limited in scope, often focusing on specific tasks, healthy tissue, or single-scale analysis. Most publicly available datasets do not simultaneously address patch-level and WSI-level evaluation, nor do they capture the morphological diversity associated with different CKD models. To facilitate an objective comparison, Table 1 summarizes representative public kidney pathology resources and pathology challenges according to species, dataset scale, annotated structures, annotation type, task definition, and condition coverage.

Table 1.

Quantitative comparison of the KPIs dataset with representative public kidney pathology resources and pathology challenges, across species, dataset scale, annotated structures, annotation type, task definition, and condition coverage. The table highlights the complementary design of KPIs relative to existing resources; datasets are grouped into kidney pathology resources and representative pathology challenges.

Dataset Species # WSIs/Images Annotated structures (count) Annotation type Task definition Conditions
Kidney pathology resources
KPIs 2024 (Ours) Preclinical mouse 50 PAS WSIs Glomeruli (>10,000) Dense pixel-level masks Patch- & WSI-level segmentation + detection Normal + 3 CKD models (multi-condition)

HuBMAP “Hacking the Kidney” (Jain et al., 2023) Clinical human 30 PAS WSIs Glomeruli (7102) Dense pixel-level masks Semantic segmentation Reference/normal (single-condition)
KPMP repositoryb (Kidney Precision Medicine Project, 2026) Clinical human Repository (no fixed count) Glomeruli, arteries, tubules (subset-dependent) Pixel masks (annotated subsets) + metadata Segmentation/classification CKD/AKI + reference (multi-condition)
NEPTUNEc (Barisoni et al., 2013, 2016) Clinical human Descriptor-scoring cohort (no fixed WSI release) Glomeruli scored (e.g., 131/315) via 51 descriptors Slide-/structure-level descriptor scores Pathology scoring/grading Glomerular diseases (multi-condition)

Representative pathology challenges
MoNuSeg 2018a (Kumar et al., 2020) Clinical human 30 train + 14 test tiles (1000 × 1000) Nuclei (~21,623 train) Instance segmentation masks Nuclei instance segmentation Multi-organ tumor tissue
CAMELYON16 (Ehteshami Bejnordi et al., 2017) Clinical human 400 WSIs Tumor metastasis regions Slide-level labels + pixel masks Metastasis detection + classification Normal vs. metastasis
CAMELYON17 (Bándi et al., 2019) Clinical human 1000 WSIs Metastases (ITC/micro/macro) Slide-level labels + subset pixel masks Patient-level pN-staging Multi-class metastasis severity
PANDA (Bulten et al., 2022b) Clinical human 10,616 WSIs (2113 patients) Gleason patterns (benign/G3/G4/G5) Slide-level ISUP grade + pixel masks Gleason grading/classification Benign + grades 1–5
a

MoNuSeg provides 1000 × 1000-pixel image tiles cropped from TCGA whole slide images, not full WSIs.

b

KPMP is a continuously growing open repository without a single canonical WSI count; reported counts are study-specific subsets.

c

NEPTUNE is primarily a descriptor-based pathologist-scoring resource rather than a dense pixel-level segmentation benchmark; reported numbers refer to glomeruli scored by readers.

Several kidney-related digital pathology resources exist, but they vary substantially in scope and are often not organized as unified, challenge-style segmentation benchmarks. For example, the HuBMAP Hacking the Kidney benchmark provides expert annotations for functional tissue units in human kidney sections (e.g., glomeruli and other structures) and has been widely used for semantic segmentation; however, it primarily targets healthy human tissue and does not emphasize multi-condition disease variability or WSI-level evaluation (Jain et al., 2023). Large clinical consortia such as KPMP (Kidney Precision Medicine Project, 2026) and NEPTUNE (Barisoni et al., 2013) have established digital pathology repositories with whole-slide kidney biopsies and rich accompanying metadata, enabling studies of renal lesion quantification and descriptor-based scoring, yet they are not consistently accompanied by dense pixel-wise annotations and standardized train/test protocols for benchmarking glomerular segmentation across diverse conditions (Kidney Precision Medicine Project, 2026; Barisoni et al., 2013). More recently, kidney pathology datasets have also been curated for nuclei-centric tasks (e.g., nuclei segmentation and cell-level assessment), which further highlights the community’s interest in kidney histology while leaving glomerulus-focused, multi-condition segmentation benchmarks comparatively underrepresented. Together, these limitations motivate a kidney-specific benchmark that combines large-scale glomerular annotations with standardized evaluation across multiple CKD conditions and both patch- and WSI-level settings. Compared with the existing publicly available resources summarized in Table 1, KPIs uniquely combines dense pixel-level glomerular annotation with standardized patch- and WSI-level segmentation and detection across multiple CKD models, highlighting its complementary role as a standardized benchmark for evaluating glomerular segmentation methods across diverse CKD models and imaging scales.

3. Challenge description

3.1. Overview

The KPIs Challenge seeks to advance glomeruli segmentation in kidney pathology using deep learning techniques. Participants are tasked with developing precise, pixel-level segmentation algorithms for glomeruli across diverse mouse models, including normal and diseased conditions. The dataset includes WSIs and patches from four mouse groups: normal, 5/6 nephrectomy (5/6Nx), diabetic nephropathy (DN), and NEP25 models, stained with PAS. The challenge emphasizes distinguishing glomeruli from surrounding tissue components under varied preparation scenarios, highlighting the algorithms’ versatility and precision.

The competition is divided into two distinct tasks, each addressing specific challenges and emphasizing different aspects of segmentation:

  • Patch-level Segmentation: This task focuses on the segmentation of glomeruli within specific image patches extracted from the WSIs. Participants must handle high-resolution data and ensure precise delineation of glomeruli under constrained spatial contexts. This task is ideal for evaluating algorithms on local-level segmentation challenges.

  • WSI-Level Segmentation: This task involves segmenting glomeruli across entire kidney slide images, which presents additional challenges such as handling large-scale data, maintaining computational efficiency, and addressing variability across the entire tissue section. It requires algorithms capable of performing robust segmentation on a global scale, adapting to diverse morphological and pathological variations.

By addressing these tasks, participants contribute to advancing automated kidney pathology image analysis, demonstrating the adaptability and effectiveness of their algorithms in both localized and comprehensive segmentation scenarios. Participants can choose to tackle either or both tasks, contributing innovative solutions to advance kidney pathology image analysis.

3.2. Data description

3.2.1. Data overview

The dataset for the KPIS Challenge includes WSIs derived from four groups of mouse models:

  • Normal group: Normal mice, sacrificed at the age of 8 weeks.

  • 5/6Nx group: Mice underwent 5/6 nephrectomy, sacrificed at 12 weeks after nephrectomy (age of 20 weeks) (Lim et al., 2014).

  • DN group: eNOS−/− db/db mice (male or female) sacrificed at the age of 18 weeks (Al Tuhaifi et al., 2024).

  • NEP25 group: Transgenic mice expressing human CD25 selectively in podocytes, sacrificed at 3 weeks after immunotoxin-induced glomerular injury (age of 11 weeks) (Lim et al., 2017).

The four mouse models included in the KPIs Challenge were selected to represent complementary and widely used experimental paradigms in CKD research. These models recapitulate distinct pathological mechanisms and glomerular morphologies, including a healthy baseline (normal), nephron-reduction-induced CKD (5/6 nephrectomy), diabetic kidney disease (diabetic nephropathy), and primary podocyte injury-induced focal segmental glomerulosclerosis (NEP25). Together, they provide a diverse preclinical testbed for evaluating segmentation robustness across heterogeneous CKD-related phenotypes.

The tissue sections were stained with PAS to visualize morphology, and the stained slides were scanned using the KF-pro-040-Hi scanner (Konfoong Bioinformation Tech Co., Ltd.) and Zeiss Axio Scan Z1 brightfield and fluorescence slide scanner (Carl Zeiss AG, Germany) at Vanderbilt University Medical Center. Each image encompasses glomeruli, tubules, interstitium, and vasculature. All glomeruli in the digital images were annotated by three experienced pathologists using QuPath software (Bankhead et al., 2017). Annotations were conducted following a unified guideline. For each image region, two pathologists independently reviewed the annotations and reached agreement; when consensus could not be reached, a third pathologist was involved to adjudicate and finalize the annotation. As the final labels were generated through a consensus-based process rather than independent parallel annotations, formal inter-annotator agreement statistics were not computed, and this is noted as a limitation of the dataset.

The WSIs are formatted in pyramid resolution levels of 1, 2, 4, 8, 16, 32, 64, formatted as TIFF files (.tiff). The image patches are provided at a fixed resolution of 2048 × 2048 pixels, formatted as PNG files (.png). This patch size was selected to balance glomerular coverage and dataset scale. In the mouse kidney, the typical glomerular diameter ranges from approximately 50–100 μm. Given this scale, smaller patches (e.g., 512 × 512) often contain only partial or very few complete glomeruli, while intermediate sizes (e.g., 1024 × 1024) substantially increase the total number of patches, leading to redundancy and higher computational burden. The chosen resolution enables each patch to include multiple complete glomeruli while maintaining a manageable number of patches for efficient training and evaluation.

WSIs were digitized using two different scanners, resulting in native scanning resolutions of 0.11 and 0.24 μm/px across samples. For benchmarking, all validation and testing images were resampled to a uniform optical magnification (40×) to ensure consistent and fair evaluation across disease models, while the original scanning resolutions are reported for transparency.

The dataset is divided into training, validation, and testing sets, with each subset containing annotations for glomeruli segmentation. Table 2 provides an overview of the dataset, detailing the number of WSIs, image resolution, and other characteristics across the subsets.

Table 2.

Summary of the KPIs Challenge dataset. Original scanning resolutions are reported for the training set. All validation and testing images were resampled to a unified optical magnification (40×) for standardized evaluation.

Property Training
Validation
Testing
Norm. 5/6Nx DN NEP25 Norm. 5/6Nx DN NEP25 Norm. 5/6Nx DN NEP25
# WSIs 5 5 5 15 2 2 2 2 3 3 3 3
# Patches 1786 648 724 1760 861 274 299 209 1166 463 391 285
Resolution (μm/px) 0.11 0.11 0.11 0.24 0.11 0.11 0.11 0.25 0.11 0.11 0.11 0.25
Optical Mag. 40× 40× 40× 20× 40× 40× 40× 40× 40× 40× 40× 40×
Digital Mag. 80× 80× 80× 40× 80× 80× 80× 40× 80× 80× 80× 40×

3.3. Challenge setup

The KPIs Challenge began with the release of the training dataset on March 30, 2024, hosted on the Synapse platform, which facilitates data sharing. The validation dataset was released on May 1, 2024, allowing participants to evaluate their algorithms locally.

The test dataset remained private, and participants were required to submit Docker containers with their algorithms. These containers were evaluated after the testing phase ended. The evaluation was performed on an Ubuntu 20.04 desktop with the following specifications:

  • CPU: Intel(R) Xeon(R) Gold 6230R CPU @ 2.10 GHz (52 threads)

  • GPU: NVIDIA RTX A6000 with 48 GB memory

  • RAM: 48 GB

  • Driver Version: 535.183.01

  • CUDA Version: 12.2

  • Docker Version: 27.0.1

The challenge concluded with the announcement of winners on September 15, 2024. Participants were invited to present their methods at the MICCAI 2024 MOVI Workshop on October 10, 2024, and encouraged to contribute to a joint manuscript for dissemination of the challenge outcomes.

3.4. Metrics and evaluation

3.4.1. Choice of metrics

The KPIs Challenge employs tailored metrics to address the specific objectives of each task:

  • Task 1: Patch-level Glomeruli Segmentation The DSC is used to evaluate the precision of segmentation at the patch level. This metric ensures accurate pixel-level overlap between predicted and ground truth masks, emphasizing the importance of precise segmentation in smaller, localized regions.

  • Task 2: WSI-level Glomeruli Segmentation and Detection The WSI-level task is designed as a single task with two complementary evaluation outputs, reflecting both segmentation quality and detection performance. WSI segmentation is significantly more challenging than patch-level segmentation due to the large image size and complexity of WSIs compared to traditional segmentation tasks. These large-scale images often require models to maintain computational efficiency while addressing the diverse spatial and morphological variations of glomeruli. To balance these challenges, the evaluation includes:

    • DSC: Measures the pixel-level accuracy of glomeruli segmentation, ensuring precise delineation.

    • F1 Score: Evaluates the balance between precision and recall, promoting comprehensive glomeruli detection. These two metrics jointly evaluate the outputs of Task 2 and are combined with equal weight to produce the final ranking, reflecting both detection comprehensiveness and segmentation precision.

WSI-level segmentation and detection are therefore evaluated as two performance metrics of the same task, rather than as separate sub-tasks.

These metrics ensure a fair and comprehensive evaluation, addressing the unique challenges posed by patch-level and WSI-level segmentation and detection tasks.

4. Participating methods

In this section, we summarize the participating methods for each challenge task, focusing on the top 3 performing teams as representatives. To keep the main text concise, we summarize the principal methodological trends and key differences here, and provide detailed per-method descriptions in the Supplementary Material (Appendices A.1 and A.2). Tables 3 and 4 provide a brief comparison of the proposed approaches, highlighting their methodology and implementation details, including the training and inference strategies. Across all top-performing submissions, ensemble learning played a central role, with teams leveraging model diversity through multiple random initializations, architectural variants, multi-scale inputs, or cross-validation folds, and combining predictions via averaging, weighted fusion, or stitching-based aggregation.

Table 3.

Summary of methods proposed by the top-performing teams for Task 1 (patch-level glomeruli segmentation). For each method, we highlight methodology and implementation details, including the training and inference strategies.

Methodology Training strategy Inference strategy
Rank Teams Methods Segmentation networks Cropped (Downsample) size Data augmentations Loss function(s) Optimization Pre-processing Ensembling Post-processing
1st place Capybara Ensembling + Stiching SegFormer-B5 768 × 768 Brightness, flips, blur. CE + DICE AdamW Cropping 3 × SegFormer-B5 Sum overlapping patches + Softmax + Crop back
2nd place agaldran Ensembling FPN/MiT 1024 × 1024 (Downsample) Custom augm.a CE + DICE Nadam Downsample + TTAb 5 × FPN/MiT Upsample
3rd place Aira Matrix Ensembling SegNeXt 1024 × 1024, 512 × 512 Custom augm.c CE + IoU loss + Entropy regularization AdamW Cropping 2 × SegNeXt Weighted average + Argmax
a

Cropping, intensity scaling, resizing, flipping, rotation, color jittering, sharpness adjustment.

b

Test Time Augmentation: Flipping, Rotation.

c

Flipping, spatial scaling, shifting, rotations, color jittering, gamma correction.

Table 4.

Summary of methods proposed by the top-performing teams for Task 2 (WSI-level glomeruli segmentation and detection). Team performance is evaluated using two complementary metrics: WSI-level Dice for segmentation quality (Task 2-Seg.) and F1-score for detection performance (Task 2-Det.).

Methodology Training strategy Inference strategy
Rank Teams Methods Segmentation networks Cropped (Downsample) size Data augmentations Loss function(s) Optimization Pre-processing Ensembling Post-processing
Seg.: 1st place (Det.: 4th place) Capybara Ensembling + Stiching SegFormer-B5 768 × 768 Brightness, flips, blur. CE + DICE AdamW Cropping 3 × SegFormer-B5 Sum overlapping patches + Softmax + Crop back
Seg.: 2nd place (Det.: 3rd place) Zhijian Life Ensembling + Stiching ResNet101 4096 × 4096 (80×)
2048 × 2048 (40×)
– – – OTSU + Cropping + Filtering noise ResNet101 Edge patch: Paste
Center patch: Crop, Paste
Seg.: 2nd place (Det.: 5th place) Deep Bio Ensembling + Stiching Hybrid.a 1024 × 1024 Custom augm.b DICE AdamW Cropping Hybrid. (4 seeds) Average + Thresholding
Det.: 1st place (Seg.: 4th place) salt_fish SAM (LoRA) + Ensembling + Stiching ViT-H (SAM) 512 × 512 Custom augm.c CE + DICE AdamW Cropping + Add extra dim. ViT-H (4 seeds) Sigmoid + Thresholding + Remove extra dim.
Det.: 2nd place (Seg.: 5th place) CVAILAB Ensembling + Stiching nnU-Net – nnU-Net augm. – – Cropping + Edge removal nnU-Net with U-Net++ and U-Net3+ Center patch: stich + resize
a

Encoder: ResNet50 + ViT-B/16; Decoder: SETR.

b

Flipping, rotations, resizing, color jittering, gaussian filtering, replacing 10% image in mini batch with ImageNet21k, Mixup, CutMix.

c

Flipping, spatial scaling, rotation, gaussian noise and filtering, contrast adjustment.

4.1. Task 1: Patch-level glomeruli segmentation

The top-performing patch-level methods shared several design characteristics, summarized in Table 3. All adopted encoder–decoder segmentation networks initialized with ImageNet-pretrained weights or pretrained backbones, spanning transformer-based and hybrid CNN–Transformer backbones: Capybara (1st) ensembled three SegFormer-B5 models, agaldran (2nd) combined an ImageNet-pretrained Feature Pyramid Network encoder with a Mix Vision Transformer decoder, and Aira Matrix (3rd) ensembled two ConvNeXt-Tiny-based SegNeXt models trained at different spatial scales. Ensemble learning was employed consistently to improve robustness—realized through cross-validation folds, disease-stratified folds, or multi-scale models—with predictions combined by simple or weighted averaging. Geometric and photometric data augmentation (e.g., flips, rotation, scaling, color jittering, and gamma adjustment) was widely used to improve generalization across heterogeneous CKD morphologies. Overall, the teams differed mainly in their choice of backbone architecture, ensemble strategy, and post-processing pipeline rather than in the underlying segmentation formulation. Detailed per-method descriptions, including training, inference, and post-processing specifics, are provided in Supplementary Section A.1.

4.2. Task 2: WSI-level glomeruli segmentation and detection

The WSI-level submissions followed a common patch-based inference framework—tissue localization, sliding-window prediction, patch-level aggregation, and post-processing—but exhibited greater methodological diversity than the patch-level task, particularly in how patch predictions were stitched into a whole-slide map (Table 4). The approaches can be grouped into segmentation-focused and detection-focused methods. Among segmentation-focused methods, Capybara (1st) reused its Task 1 SegFormer-B5 ensemble, Zhijian Life employed a UPerNet–ResNet101 pipeline with multi-magnification (80×/40×) patch extraction, and Deep Bio used a hybrid ResNet50–ViT-B encoder with a SETR decoder together with strong augmentation, including MixUp and CutMix; most relied on multi-seed or multi-model ensembling to improve robustness. Among detection-focused methods, salt_fish fine-tuned the Segment Anything Model (SAM, ViT-H) with LoRA, highlighting the potential of foundation models for glomerular localization, while CVAILAB used an nnU-Net framework with U-Net++ and U-Net3+ variants and adaptive, edge-aware cropping. Two dominant patch-aggregation strategies emerged: (1) probability-map fusion followed by normalization or thresholding, and (2) crop-and-paste stitching that retains only the central region of overlapping predictions to reduce boundary artifacts. The relative emphasis on segmentation precision versus detection recall is reflected in the complementary Dice and F1 rankings. Detailed per-method implementations are provided in Supplementary Section A.2.

5. Results

Participants were required to submit their inference models in the form of a Docker container by August 1, 2024. These submissions were evaluated using our withheld test set, and the final results and leaderboard for the challenge were announced on September 15, 2024. A total of 24 valid submissions were received for Task 1 (Patch-level segmentation) and 15 valid submissions for Task 2 (WSI-level segmentation). The complete leaderboards for all participants can be found at.3

This section presents the overall segmentation performance for Task 1 and Task 2. Specifically, it includes both qualitative and quantitative results for the top 10 teams on the leaderboard for each task. Reported mean and standard deviation values are computed across test samples (patches or WSIs) for each submission and do not reflect variability across multiple model initializations or repeated inference runs. The complete benchmarking details, derived from an open-source toolkit for analyzing and visualizing challenges (Wiesenfarth et al., 2021), are provided in Section Supplementary data (Appendix B).

5.1. Patch-level segmentation performance

This section reports the top 10 teams on the patch-level glomeruli segmentation leaderboard. Table 5 presents their ranked Dice scores with mean ± standard deviation (SD), while Fig. 5a visually illustrates the Dice score distribution as a boxplot. Both the table and figure show that the top 10 teams performed well, with closely aligned mean and median Dice scores, indicating high consistency across their models. Capybara leads with a Dice score of 94.51 ± 6.71, followed by agaldran and Aira Matrix at 94.28 ± 5.81 and 94.28 ± 5.89, respectively. These top 3 teams, summarized as representatives in the previous section, exhibit relatively high mean Dice scores and lower standard deviations compared to lower-ranked teams, which show greater variability, as indicated by SD values exceeding 7.24. However, standard deviation should be interpreted as a measure of stability rather than ranking, as some mid-ranked teams exhibit low variance but consistently lower mean performance, while higher-ranked methods may show slightly increased variability on more challenging samples. The narrow performance gap among the top-ranked methods suggests a near-saturation regime for patch-level evaluation, where performance differences may be of limited practical significance. Fig. 2a confirms this trend, with most teams demonstrating small variability, while others show more inconsistency. Fig. 3 provides qualitative patch-level segmentation predictions alongside ground truth, demonstrating stable model performance across all listed teams for four types of glomeruli samples. Top-performing teams (left columns) handle boundary segmentation more effectively than lower-ranked teams (right columns), as seen in the 5/6Nx and NEP25 rows. A challenge in this task is the large size of the given image patches (2048 × 2048), where patch cropping could compromise accuracy at the boundaries, particularly when merging predictions. The stronger performance of the top teams (e.g., Capybara and Aira Matrix) suggests they have implemented effective post-processing strategies to address this challenge.

Table 5.

Task 1: Patch-level diseased glomeruli segmentation leaderboard. The submissions of the top 10 teams are presented, along with their Dice scores, shown as Mean ± SD.

Rank Team Dice score
1 Capybara 94.51 ± 6.71
2 agaldran 94.28 ± 5.81
3 Aira Matrix 94.28 ± 5.89
4 mafi95 94.16 ± 6.86
5 Deep Bio 93.64 ± 6.00
6 Birmingham Vanguard AI Lab 93.62 ± 4.52
7 p53 93.23 ± 7.65
8 Bio-Totem 93.08 ± 7.24
9 CVAILAB 92.79 ± 8.28
10 NJUST Vision Group 92.70 ± 7.73

Fig. 5.

Fig. 5.

Task performance across disease models at patch- and WSI-level. - Boxplots summarize segmentation performance grouped by disease model (5/6 nephrectomy, diabetic nephropathy, NEP25, and normal) for both Task 1 (patch-level Dice) and Task 2 (WSI-level Dice and F1-score). Patch-level results exhibit consistently high performance with limited variability across disease models, indicating stable local segmentation accuracy. In contrast, WSI-level performance shows substantially greater variability, particularly for severe disease models, reflecting the increased difficulty of slide-level detection and segmentation under heterogeneous tissue conditions. Dashed lines indicate overall mean and median performance across teams.

Fig. 2.

Fig. 2.

Box plots of quantitative results on different tasks from participants. This figure shows the box plots of the quantitative performance of participants on two tasks. Results from the top 10 teams are reported. DSC scores are provided for Task 1, while both DSC scores and F1 scores are presented for Task 2.

Fig. 3.

Fig. 3.

Patch-level quantitative results - The segmentation results from different participants’ methods on glomeruli under various experimental conditions (normal, NEP25, DN, and 5/6Nx).

5.2. WSI-level segmentation performance

The transition from patch-level to slide-level analysis introduced new challenges, particularly in handling large-scale gigapixel images, which demand significant computational resources. This necessitates image cropping to manage these vast images, followed by addressing boundary effects when merging predictions from cropped patches back to the WSI level. In this section, we present the WSI-level glomeruli instance segmentation and detection results for Task 2, evaluated using Dice and F1 scores.

5.2.1. Instance segmentation

We also report the top 10 teams on the WSI-level glomeruli instance segmentation leaderboard (Table 6). Both Dice and F1 scores are used to calculate the overall rank for this task. Fig. 2b and c visually depict the distributions of Dice and F1 scores as boxplots. Fig. 4 shows a zoomed-in visualization of WSI-level segmentation masks alongside ground truth, where the top-performing teams are ranked from left to right across the columns.

Table 6.

Task 2a: WSI-level diseased glomeruli instance segmentation leaderboard. The top 10 teams are provided, with the overall rank based on the F1 score (detection) and Dice score (segmentation) performance. The values for both metrics are presented as Mean ± SD.

Overall rank Team Dice score F1 score Dice rank F1 rank
1 Capybara 94.64 ± 0.89 88.63 ± 5.80 1 4
2 Zhijian Life 93.36 ± 1.99 89.99 ± 4.31 4 3
2 Deep Bio 94.48 ± 1.28 87.96 ± 6.08 2 5
4 salt_fish 89.81 ± 2.18 91.33 ± 3.84 8 1
5 CVAILAB 89.37 ± 3.45 90.88 ± 4.35 9 2
5 p53 92.74 ± 1.62 86.81 ± 5.42 5 6
7 BozekLab 91.35 ± 3.05 86.48 ± 7.96 6 7
7 mafi95 93.68 ± 1.14 74.30 ± 17.18 3 10
9 TeamTiger 89.83 ± 5.26 84.96 ± 7.81 7 9
10 XJTU 84.47 ± 5.90 85.84 ± 7.14 10 8
Fig. 4.

Fig. 4.

WSI-level quantitative results - This figure shows the segmentation results from different participants’ methods on glomeruli under various experimental conditions (normal, NEP25, DN, and 5/6Nx). The yellow bounding box highlights the zoomed-in region on the WSI, with the zoomed-in images below showing the corresponding area.

Overall, compared to the patch-level task results from Table 6 and Fig. 2a, the WSI-level instance segmentation task is more challenging, as indicated by the broader range of both segmentation (Dice) and detection (F1) metrics in the boxplots. Capybara emerges as the top performer, leading in Dice with a score of 94.64 ± 0.89 and ranking 4th in F1 with a score of 88.63 ± 5.80. Zhijian Life and Deep Bio follow closely, sharing the second rank, with Zhijian Life excelling in detection (F1) and Deep Bio leading in segmentation (Dice). The top teams consistently demonstrate higher mean Dice and F1 scores, with relatively smaller SD compared to lower-ranked teams. Among all teams, Capybara has the highest mean Dice score and the lowest SD of 0.89, making it the only team with a Dice SD below 1. Zhijian Life achieves a Dice score of 93.36 ± 1.99 and an F1 score of 89.99 ± 4.31, while Deep Bio scores 94.48 ± 1.28 in Dice and 87.96 ± 6.08 in F1. These smaller SD values indicate that the top teams’ models are more stable, with less variation in their performance across different WSI samples. In contrast, lower-ranked teams show larger SD values, with Dice SD exceeding 5 and F1 SD reaching as high as over 17. For both evaluation metrics, the difference between the highest and lowest mean scores can exceed 10, highlighting the inconsistency and variability in WSI-level segmentation and detection tasks. To visually confirm, an example of the zoomed-in visualization of a WSI-level segmentation mask in Fig. 4 shows the segmentation performance among different participants. As shown, for the glomeruli types NEP25, DN, and 5/6Nx, the difference in segmentation accuracy is more pronounced between top-performing (left columns) and lower-ranked teams (right columns).

5.2.2. Detection

For Task 2, detection performance is quantified using the F1 score (mean ± SD) computed from the instance-level predictions. Table 7 reports the top 10 teams on the WSI-level glomeruli detection leaderboard, ranked by F1 scores with mean ± SD. Fig. 2c shows the F1 score distribution as a boxplot. Similarly, Fig. 4 visually compares the WSI-level glomeruli detection performance with zoomed-in regions, where the top-performing teams are ranked from left to right across the columns.

Table 7.

Task 2b: WSI-level diseased glomeruli detection leaderboard. The submissions of the top 10 teams are presented, along with their F1 scores, shown as Mean ± SD.

Rank Team F1 score
1 salt_fish 91.33 ± 3.84
2 CVAILAB 90.88 ± 4.35
3 Zhijian Life 89.99 ± 4.31
4 Capybara 88.63 ± 5.80
5 Deep Bio 87.96 ± 6.08
6 p53 86.81 ± 5.42
7 BozekLab 86.48 ± 7.96
8 XJTU 85.84 ± 7.14
9 TeamTiger 84.96 ± 7.81
10 mafi95 74.30 ± 17.18

Among the participants, salt_fish emerges as the top performer, achieving an F1 score of 91.33 ± 3.84. CVAILAB and Zhijian Life follow closely, with F1 scores of 90.88 ± 4.35 and 89.99 ± 4.31, respectively. As observed earlier, the WSI-level task proposes new challenges and leads to more varied team performance. However, the top-performing teams in detection tasks generally have better rank in instance segmentation Task 2, which can be also derived from Tables 3 and 4. Similarly, these teams exhibit high F1 detection scores with relatively low variability in their results, as indicated by smaller SD. In contrast, teams ranked lower demonstrate more variability, with SD exceeding 7 for most, which indicates the importance of using rigorous pre-processing and post-processing methods (e.g., Zhijian Life, CVAILAB), ensembling methods, or even foundation models (e.g., salt_fish) in the slide-wise task.

5.3. Per-disease model performance and failure case analysis

To better understand performance variability beyond aggregate metrics, we analyze both patch-level and WSI-level results grouped by disease model (Fig. 5). At the patch level, Dice scores remain consistently high across all CKD conditions, with limited inter-team variability, indicating that most methods reliably capture local glomerular features regardless of disease type. This stability suggests that patchlevel evaluation primarily reflects local appearance cues and may be insufficient to fully differentiate top-performing methods. In contrast, WSI-level performance exhibits substantially greater variability across disease models, as reflected in both Dice and F1-score distributions. Performance degradation is more pronounced for severe disease conditions, such as NEP25 and diabetic nephropathy, where morphological heterogeneity, disrupted tissue structure, and atypical glomerular appearances are more common.

These results highlight that slide-level detection and segmentation are more sensitive to disease severity and tissue context, and are strongly influenced by post-processing strategies and global reasoning capabilities. Qualitative inspection of the WSI-level predictions further reveals three recurring categories of error: false-negative detections (missed glomeruli), false-positive detections (spurious predictions), and boundary inaccuracies, defined as detected glomeruli whose intersection-over-union with the ground truth falls below 0.5. Representative examples across the four CKD models are provided in Fig. 6, where these error types are color-coded directly on each team’s prediction. Errors are comparatively infrequent in the normal model and become more common and more varied in the diseased models, consistent with the per-disease trends reported above. Together, these analyses demonstrate that while patch-level segmentation performance is largely saturated, WSI-level evaluation provides a more discriminative and clinically relevant assessment of model robustness.

Fig. 6.

Fig. 6.

Representative failure cases of participating methods on the WSI-level task. Representative examples illustrating common glomerular segmentation errors across the four CKD models (normal, NEP25, DN, and 5/6Nx). From left to right: the original PAS image, the manual (ground-truth) annotation, and the predictions of the top-performing teams. Segmentation errors are color-coded directly on each prediction: false negatives in red (missed glomeruli), false positives in blue (spurious predictions), and boundary inaccuracies in yellow (detected glomeruli whose intersection-over-union with the ground truth is below 0.5); correctly segmented regions are shown in white. These examples provide visual evidence supporting the qualitative failure analysis in Section 5.3 and illustrate the challenges of robust glomerular segmentation under diverse pathological conditions.

5.4. Computational efficiency analysis

Processing WSIs is computationally demanding, and efficiency is an important consideration for practical deployment. To address this, we evaluated the computational performance of participating methods under standardized benchmarking conditions.

All submitted Docker containers were executed on a unified evaluation server equipped with an NVIDIA A6000 GPU (48 GB memory). Inference time and peak GPU memory usage were recorded during evaluation without modifying the submitted code.

Task 1 (Patch-level).

Table 8 reports the inference time per 350 test patches and peak GPU memory usage for the top-performing teams. Patch-level inference is generally efficient, with relatively modest memory consumption across methods. Differences in runtime are primarily influenced by model size and ensemble strategies.

Table 8.

Task 1: Patch-level computational efficiency (NVIDIA A6000 48 GB).

Rank Team Time (s) GPU-Mem (MiB)
1 Capybara 2192.06 992
2 agaldran 393.00 718
3 Aira Matrix 399.04 896
4 mafi95 2560.94 788
5 Deep Bio 243.91 9202
6 Birmingham Vanguard AI Lab 848.73 954
7 p53 204.25 940
8 Bio-Totem 1020.99 986
9 CVAILAB 156.75 954
10 NJUST Vision Group 150.00 902

Task 2 (WSI-level).

Table 9 summarizes the inference time per WSI and peak GPU memory usage. Compared to patch-level processing, WSI-level inference requires substantially greater computational resources due to sliding-window strategies, overlapping patch aggregation, and ensemble prediction. Methods employing larger backbones or multiple ensemble members generally exhibit higher memory usage and longer inference times.

Table 9.

Task 2: WSI-level computational efficiency (NVIDIA A6000 48 GB).

Rank Team Time (min) GPU-Mem (MiB)
1 Capybara 34.67 972
2 Zhijian Life 141.20 964
2 Deep Bio 9.66 9815
4 salt_fish 60.35 9240
5 CVAILAB 2.35 994
5 p53 29.17 832
7 BozekLab 10.61 858
7 mafi95 262.58 7382
9 TeamTiger 1.17 7594
10 XJTU 277.98 8632

Overall, these results highlight the trade-off between predictive performance and computational cost, particularly at the WSI level. While several top-performing approaches achieve strong segmentation and detection accuracy, their inference latency and memory requirements should be considered in real-world deployment scenarios.

6. Discussion

The performance of the leading participants in Task 1 was remarkably similar, with Dice scores exceeding 94% and standard deviations below 7, indicating that most methods performed at a high level with only subtle differences in segmentation accuracy (Table 5, Fig. 5). This clustering of top-performing results indicates a near-saturation regime for patch-level evaluation, where performance differences among leading methods may be of limited practical significance. However, in Task 2, greater variation in performance was observed across teams, as reflected in the broader range of Dice and F1 scores (Table 6, Table 7, Fig. 2.b and c). This variability suggests that while patch-based segmentation models generalize well, WSI-level detection and segmentation are significantly influenced by ensembling strategies and post-processing techniques, underscoring the practical differences in the robustness and adaptability of leading models.

The results of Task 1 demonstrate strong segmentation performance across participants, with subtle variations among the top-performing methods. The consistently high Dice scores suggest that the dataset and task were effective in isolating patch-level features critical to glomerular segmentation. The leading methods, such as Capybara and Aira Matrix, leveraged ensemble techniques and multi-scale processing, which underscores the importance of integrating local patch-level detail with robust model ensembling. At the same time, these results reinforce the observation that patch-level evaluation alone may be insufficient to clearly differentiate state-of-the-art approaches. Future iterations of this challenge might benefit from introducing more variability in patch conditions to further differentiate model capabilities, particularly in handling atypical or edge-case glomeruli.

The transition from patch-level to slide-level analysis presented new challenges, as evidenced by the broader range of performance metrics, particularly in F1 scores. The variability in slide-wise detection highlights the importance of ensembling and post-processing strategies to maintain precision and recall at larger scales. While the Capybara model achieved the highest segmentation scores, other participants demonstrated competitive detection accuracy, reflecting a trade-off between segmentation precision and detection comprehensiveness. The inclusion of large-scale data handling and diverse morphological conditions in Task 2 provided a robust testbed for evaluating scalability and adaptability.

A recurring trend among high-performing methods was the use of transformers or hybrid architectures, combining CNNs with transformers, for constructing the image encoder. This approach effectively captured both local and global features, a critical requirement for tasks spanning diverse spatial scales. The application of TTA and advanced data augmentation techniques, such as CutMix and MixUp, further enhanced model robustness, suggesting that these strategies are indispensable for generalizing across heterogeneous and imbalanced datasets. Additionally, to handle boundary effects while merging patchwise predictions back into the slide-wise mask, model ensembling is used along with stitching. There are two main categories of stitching methods. The first involves merging patches straightforwardly and then applying normalization or thresholding to refine the merged masks (Capybara, Aira Matrix, Deep Bio, and salt_fish). The second is the crop-and-paste method, which, depending on the data and task, handles image patches at the edges of glomeruli tissue or at the center of the glomeruli tissue differently (e.g., Zhijian Life and CAVILAB).

Comparative Analysis of Top-performing Methods.

A closer comparison of the top-performing approaches reveals that performance differences are primarily driven by architectural choices, ensembling strategies, and post-processing robustness rather than patch-level segmentation accuracy alone. The winning method, Capybara, consistently achieved strong performance across both segmentation (Dice) and detection (F1) metrics by combining a high-capacity transformerbased backbone (SegFormer-B5), multi-model ensembling, and a robust stitching strategy that effectively mitigates boundary artifacts. This combination resulted in not only high mean performance but also the lowest variance across WSIs, indicating strong robustness to diverse tissue conditions.

In contrast, Zhijian Life and Deep Bio achieved competitive performance through complementary strengths. Zhijian Life emphasized multi-scale processing and explicit handling of edge and center patches, which improved detection sensitivity but introduced greater variability in segmentation consistency. Deep Bio leveraged a hybrid CNN–transformer architecture with strong data augmentation and ensembling, achieving high segmentation accuracy but comparatively weaker detection performance in challenging WSIs. The detection-focused approach salt_fish demonstrated that foundation models with LoRA-based fine-tuning can excel at object-level detection, but may require additional refinement to achieve comparable boundary precision in dense segmentation tasks.

Overall, these comparisons indicate that the leading performance in Task 2 arises from a balanced integration of global context modeling, dense ensembling, and carefully designed post-processing, whereas methods optimizing primarily for either segmentation precision or detection recall tend to exhibit trade-offs under heterogeneous WSI conditions.

Foundation Models in Kidney Pathology.

Recent advances in foundation models have begun to influence computational pathology, offering strong general-purpose representations learned from large-scale, heterogeneous datasets. In this challenge, most participating teams adopted ImageNet-pretrained architectures due to their maturity, efficiency, and ease of adaptation to dense prediction tasks. However, foundation models were explicitly explored by some teams, most notably the top-performing WSI-level detection method (salt_fish), which employed the SAM with LoRA fine-tuning. This approach demonstrated strong detection performance, highlighting the potential of foundation models for glomerular localization. At the same time, the reliance on patch-based inference, memory constraints, and the need for domain-specific adaptation remain key challenges for deploying large foundation models in gigapixel WSI segmentation tasks. These observations suggest that foundation models represent a promising direction for future kidney pathology benchmarks, particularly when combined with efficient adaptation and domain-aware post-processing strategies.

Limitations and Future Directions:

While the dataset included a variety of CKD models, future challenges could benefit from incorporating more extensive interspecies data and additional staining techniques to simulate broader clinical scenarios. In particular, the exclusive use of preclinical rodent models in the current challenge limits direct clinical generalizability, and future benchmarks should incorporate human kidney pathology data to explicitly evaluate cross-species robustness and translational performance. Such cross-species evaluation would enable systematic assessment of domain shift between rodent and human kidney pathology and help identify model components that generalize across species. Expanding the dataset to include both normal tissue objects and lesion objects could enhance the clinical value of the models. Currently, slide-wise glomeruli detection and segmentation rely heavily on pretrained weights and ensembling, which overlook global morphology and multi-scale patterns inherent to the unique data type of pathology. Encouraging solutions with fewer preprocessing requirements and incorporating designs that leverage the unique properties of pathological data could help address these limitations.

7. Conclusion

The KPIs Challenge was organized to address existing limitations in glomerular segmentation research by expanding the dataset to include whole kidney sections from diseased rodent models. This competition aimed to advance state-of-the-art segmentation methods for glomerular identification across diverse CKD models, challenging participants to develop robust algorithms capable of pixel-level segmentation under varying tissue conditions and preparation scenarios. By addressing challenges such as variations in glomerular size, shape, and structural integrity, the KPIs Challenge not only fostered innovation but also demonstrated the adaptability and precision of modern segmentation approaches, paving the way for improved tools in kidney pathology research.

Supplementary Material

supplement 2
supplement 1

Acknowledgments

This research was supported by NIH R01DK135597 (Huo), DoD HT9425-23-1-0003 (HCY), NSF 2434229 (Huo), and KPMP Glue Grant. This work was also supported by the WCM Radiology AIMI Fellowship, Vanderbilt Seed Success Grant, Vanderbilt Discovery Grant, and VISE Seed Grant. This project was supported by The Leona M. and Harry B. Helmsley Charitable Trust grant G-1903-03793 and G-2103-05128. We extend gratitude to NVIDIA for their support by means of the NVIDIA hardware grant. This work was also supported by NSF NAIRR Pilot Award NAIRR240055.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Yuankai Huo reports equipment, drugs, or supplies was provided by NVIDIA Corp. Yuankai Huo reports financial support was provided by National Institutes of Health. Haichun Yang reports financial support was provided by National Institutes of Health. Bennett Landman reports financial support was provided by National Institutes of Health. Yuankai Huo reports financial support was provided by National Science Foundation. Bennett Landman reports financial support was provided by National Science Foundation. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary materials

A.1. Detailed descriptions of participating methods — task 1 (patch-level segmentation)

(The following per-method descriptions were relocated from main-text Section 4.1 in response to Reviewer 5, Comment 3.)

Capybara (1st place, Cap) (Cap, 2024).

Model architecture:

The proposed glomerulus segmentation pipeline is based on patch-level model ensembling and stitching. This proposed pipeline uses three SegFormer-B5 models (Xie et al., 2021) with a crop size of 768 × 768 during both training and inference.

Training strategy:

To account for the diverse CKD conditions in the challenge dataset, the training data was stratified into three folds based on disease model and WSI name, and one SegFormer-B5 model was trained per fold.

Ensemble strategy:

The three SegFormer-B5 models, each trained on a different fold, form an ensemble. Their patch-level predictions are averaged to obtain the final prediction for each patch before stitching.

Inference procedure:

For each 2048 × 2048 test patch, sliding-window inference was performed using a 768 × 768 window and a stride of 576 (75% of the window size), and predictions from the three models were averaged.

Post-processing and stitching:

Overlapping patches were stitched using patch coordinates and WSI identifiers; in intersecting regions, raw prediction values were summed. The stitched prediction map for each WSI was normalized using the softmax function, and the final output was cropped back to 2048 × 2048 for submission.

Code is available at https://github.com/huuquan1994/wsi_glomerulus_seg.

agaldran (2nd place, Galdran et al.) (Sánchez et al., 2024).

Model architecture:

The method performs glomerulus segmentation on smaller image patches of the WSI. The input image patches were resized to 1024 × 1024 for both model training and inference. The segmentation model follows an CNN–ViT architecture, using an ImageNet-pretrained Feature Pyramid Network (FPN) (Lin et al., 2017) as the encoder and a Mix Vision Transformer (MiT) (Xie et al., 2021) as the decoder. The FPN extracts multi-scale features through a top-down pathway with lateral connections, while the transformer-based MiT decoder integrates these features to produce the final segmentation map, preserving both global and local context.

Training strategy:

The model was trained using a five-fold cross-validation scheme. The final prediction is obtained by ensembling five cross-validation models via mean aggregation.

Ensemble strategy:

The predictions from the five models obtained via five-fold cross-validation are averaged. This reduces bias caused by random training–validation splits and improves robustness.

Inference procedure:

During inference, Test-Time Augmentation (TTA) was applied by flipping each patch horizontally and/or vertically. The final segmentation result was obtained by averaging predictions from the five cross-validation models.

Post-processing:

The 1024 × 1024 predictions are then resized to 2048 × 2048 for submission.

Code is available at https://github.com/agaldran/kpis/tree/main.

Aira Matrix (3rd place, A. Kumar et al.) (Kumar et al., 2024).

Model architecture:

This method uses an ensemble SegNeXt-based pipeline (Guo et al., 2022) for patch-level glomerulus segmentation. This ensemble combines two SegNeXt (Guo et al., 2022) models trained at different spatial scales. Specifically, two separate SegNeXt models are trained using randomly cropped patches of size 512 × 512 and 1024 × 1024, respectively, and both SegNeXt models use ConvNeXt-Tiny (Liu et al., 2022) as the backbone.

Training strategy:

The encoder network is initialized using ImageNet-pretrained weights. The method introduces an Entropy Regularized Segmentation (ERS) loss, which combines cross-entropy loss and IoU loss with an additional entropy regularization term.

Ensemble strategy:

The prediction masks from the two SegNeXt models, trained at different spatial scales, are merged using a weighted averaging strategy, thereby enhancing the robustness of the method.

Inference procedure:

During inference, the two SegNeXt models are applied to preprocessed image patches at their corresponding input sizes (512 × 512 and 1024 × 1024). The 512 × 512 predictions learn more fine local details and 1024 × 1024 predictions captures more global context.

Post-processing:

The two outputs are merged using a weighted average, and the final prediction is obtained by applying the Argmax function.

A.2. Detailed descriptions of participating methods — task 2 (wsi-level segmentation and detection)

(The following per-method descriptions were relocated from main-text Section 4.2 in response to Reviewer 5, Comment 3.)

A.2.1. Segmentation-focused WSI-level approaches

Capybara (1st place, Cap) (Cap, 2024).
Model architecture:

The method uses the same patch-level ensemble segmentation pipeline as in Task 1.

Training strategy:

The training setup is identical to Task 1. The training data was stratified into three folds based on disease model and WSI name, and one SegFormer-B5 model was trained per fold.

Ensemble strategy:

The patch-level ensemble is identical to Task 1. The three SegFormer-B5 models, each trained on a different fold, are ensembled by averaging their predictions.

Inference procedure:

For WSI-level inference, a sliding-window strategy was applied with a crop size of 2048 × 2048 and a stride size of 1024 × 1024 to extract patches from each WSI. The same ensemble model from Task 1 was applied to generate patch-level predictions.

Post-processing and stitching:

Patch-level predictions were merged into a unified WSI-level prediction map using the proposed stitching strategy. For overlapping patches, the raw prediction values in intersecting regions were summed. The final stitched prediction map was normalized using the softmax function, which improves detection coverage, particularly for glomeruli located near patch borders.

Code is available at https://github.com/huuquan1994/wsi_glomerulus_seg.

Zhijian Life (2nd place, Cai et al.) (Cai et al., 2024).
Model architecture:

The method uses a WSI-level diseased glomeruli segmentation pipeline based on the UPerNet (Xiao et al., 2018) framework, with ResNet101 (He et al., 2016) backbone network.

Training strategy:

Before model training, OTSU thresholding is applied to identify tissue regions, and small areas (fewer than 2000 pixels) are removed to reduce noise. For training data preparation, both WSIs and their corresponding masks are processed using an overlapping patch extraction strategy at two magnifications: 4096 × 4096 (80×) and 2048 × 2048 (40×). The extracted patches are then used to train UPerNet segmentation model with ResNet101 backbone.

Ensemble strategy:

N/A

Inference procedure:

During inference, patch-level predictions are aggregated to generate the final WSI-level segmentation mask. An allblack image (1D717) with the same resolution as the WSI is first initialized.

Post-processing and stitching:

Two patch aggregation strategies are then applied: if a patch lies on the tissue boundary (top, bottom, left, or right), the full patch-level prediction is directly pasted into ϑ; otherwise, only the central region of the patch prediction is cropped and pasted into ϑ.

Code is available at https://github.com/ZhiJianLife/KPIs2024_ZhiJianLife_task2

Deep Bio (2nd place, Lee et al.).
Model architecture:

The method adopts a hybrid segmentation architecture to improve performance under limited training data. The encoder combines low-level features extracted by ResNet50 (He et al., 2016) with the global attention mechanism of ViT-B (Dosovitskiy, 2020). The segmentation decoder is based on the SETR (Zheng et al., 2021) segmentation head. WSIs are tiled into 1024 × 1024 patches with 50% overlap horizontally and vertically for both training and inference.

Training strategy:

Both the CNN and transformer components are initialized with ImageNet-pretrained weights. To mitigate insufficient training data, strong data augmentation is applied. In addition to standard augmentations, 10% of images in each mini-batch are replaced with images from ImageNet-21K. Advanced augmentations, including MixUp (Zhang, 2017) and CutMix (Yun et al., 2019), are also used.

Ensemble strategy:

To enhance robustness, four models trained with different random seeds are ensembled by averaging their heatmaps to produce the patch-level segmentation predictions.

Inference procedure:

During inference, patch-level segmentation outputs are merged into a WSI-level prediction map by averaging values in overlapping regions.

Post-processing and stitching:

Heatmaps from the four ensemble models are averaged, and a threshold of 0.5 is applied to produce the final segmentation mask.

A.2.2. Detection-focused wsi-level approaches

salt_fish (1st place, Chen et al.) (Chen et al., 2024).
Model architecture:

The method uses the Segment Anything Model (SAM) (Kirillov et al., 2023) as the image encoder and applies Low-Rank Adaptation (LoRA) (Hu et al., 2021) for fine-tuning. Four SAM (ViT-H) models with different initializations are trained to form an ensemble.

Training strategy:

The input images are normalized, and an additional dimension is added to the 2D labels. During training, the input data are randomly cropped into 512 × 512 patches. SAM (ViT-H) is fine-tuned using LoRA with four different initializations.

Ensemble strategy:

The ensemble leverages different initialization seeds prior to fine-tuning. Four SAM (ViT-H) models are fine-tuned with different random seeds, and the patch-level predictions are obtained by averaging the outputs of the four models.

Inference procedure:

During inference, sliding-window prediction is performed with a patch size of 512 × 512. The model outputs are passed through a sigmoid activation function and thresholded at 0.5.

Post-processing and stitching:

The added label dimension is then removed, and objects smaller than 10,000 pixels are filtered out. An ensemble of four LoRA-finetuned SAM models is used, with predictions merged after thresholding. Finally, patch-level segmentation outputs are merged into a WSI-level prediction.

Code is available at https://github.com/jj-ccc/KPIs2024

CVAILAB (2nd place, Hou et al.) (Hou et al., 2024).
Model architecture:

The method uses an nnU-Net (Isensee et al., 2021) segmentation framework with two modified U-Net (Ronneberger et al., 2015) variants, U-Net++ (Zhou et al., 2018) and U-Net3+ (Huang et al., 2020), which introduce denser connections to improve feature reuse and generalizability.

Training strategy:

Training patches are extracted with overlaps, and an edge removal strategy is applied to prevent glomeruli fragmentation near patch boundaries. An adaptive cropping strategy is integrated into the pipeline to reduce edge effects during patch-based segmentation.

Ensemble strategy:

Multiple nnU-Net variants (U-Net++ and U-Net3+) are trained to improve robustness, and their outputs are averaged to produce the final patch-level predictions.

Inference procedure:

During inference, overlapping patches are generated using adaptive cropping, and predictions from multiple nnU-Net variants are ensembled to produce more robust segmentation masks.

Post-processing and stitching:

Only the central regions of patch predictions are stitched back into the WSI-level mask using recorded patch coordinates.

Code is available at https://github.com/jasminethurder/KPI2024

Zhijian Life (3rd place, Cai et al.) (Cai et al., 2024).
Model architecture:

The method uses a WSI-level diseased glomeruli segmentation pipeline based on the UPerNet (Xiao et al., 2018) framework, with ResNet101 (He et al., 2016) backbone network.

Training strategy:

Before model training, OTSU thresholding is applied to identify tissue regions, and small areas (fewer than 2000 pixels) are removed to reduce noise. For training data preparation, both WSIs and their corresponding masks are processed using an overlapping patch extraction strategy at two magnifications: 4096 × 4096 (80×) and 2048 × 2048 (40×). The extracted patches are then used to train UPerNet segmentation model with ResNet101 backbone.

Inference procedure:

During inference, patch-level predictions are aggregated to generate the final WSI-level segmentation mask. An allblack image (ϑ) with the same resolution as the WSI is first initialized.

Ensemble strategy:

N/A

Post-processing and stitching:

Two patch aggregation strategies are then applied: if a patch lies on the tissue boundary (top, bottom, left, or right), the full patch-level prediction is directly pasted into ϑ; otherwise, only the central region of the patch prediction is cropped and pasted into ϑ.

Code is available at https://github.com/ZhiJianLife/KPIs2024_ZhiJianLife_task2

A.3. Benchmarking analysis of diseased glomeruli segmentation across two tasks

The supplementary materials provide an in-depth benchmarking analysis of diseased glomeruli segmentation across two tasks. Task 1 focuses on patch-level diseased glomeruli segmentation, assessing ten algorithms over 2305 cases. The ranking methodology relies on the mean metric values, and multiple visualization techniques, such as boxplots, podium plots, and ranking heatmaps, illustrate performance differences among algorithms. Stability analyses using bootstrap sampling, significance maps, and ranking robustness comparisons further validate the results. Task 2 extends the evaluation to WSI-level diseased glomeruli instance segmentation, comparing algorithm performance across different ranking criteria, including Dice and F1 scores. This study examines 10 algorithms over 12 cases and integrates cross-task insights, exploring algorithm variability through dendrogram clustering and ranking robustness across different evaluation methods. These supplementary results offer valuable insights into algorithmic performance, reliability, and comparative effectiveness in glomeruli segmentation tasks.

Appendix B. Supplementary data

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.media.2026.104234.

Footnotes

CRediT authorship contribution statement

Ruining Deng: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Tianyuan Yao: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Yucheng Tang: Validation, Supervision, Software, Resources, Funding acquisition, Data curation. Junlin Guo: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Formal analysis. Siqi Lu: Writing – review & editing, Writing – original draft, Visualization, Methodology, Investigation. Juming Xiong: Writing – review & editing, Writing – original draft, Visualization, Formal analysis. Lining Yu: Validation, Software. Quan Huu Cap: Writing – review & editing, Software, Methodology, Investigation. Pengzhou Cai: Writing – review & editing, Software, Methodology, Investigation. Libin Lan: Writing – review & editing, Software, Methodology, Investigation. Ze Zhao: Writing – review & editing, Software, Methodology, Investigation. Adrian Galdran: Writing – review & editing, Software, Methodology, Investigation. Amit Kumar: Writing – review & editing, Software, Methodology, Investigation. Gunjan Deotale: Writing – review & editing, Software, Methodology, Investigation. Dev Kumar Das: Writing – review & editing, Software, Methodology, Investigation. Inyoung Paik: Writing – review & editing, Software, Methodology, Investigation. Joonho Lee: Writing – review & editing, Software, Methodology, Investigation. Geongyu Lee: Writing – review & editing, Software, Methodology, Investigation. Yujia Chen: Writing – review & editing, Software, Methodology, Investigation. Wangkai Li: Writing – review & editing, Software, Methodology, Investigation. Zhaoyang Li: Writing – review & editing, Software, Methodology, Investigation. Xuege Hou: Writing – review & editing, Software, Methodology, Investigation. Zeyuan Wu: Writing – review & editing, Software, Methodology, Investigation. Shengjin Wang: Writing – review & editing, Software, Methodology, Investigation. Maximilian Fischer: Writing – review & editing, Software, Methodology, Investigation. Lars Krämer: Writing – review & editing, Software, Methodology, Investigation. Anghong Du: Writing – review & editing, Software, Methodology, Investigation. Le Zhang: Writing – review & editing, Software, Methodology, Investigation. Maria Sánchez Sánchez: Writing – review & editing, Software, Methodology, Investigation. Helena Sánchez Ulloa: Writing – review & editing, Software, Methodology, Investigation. David Ribalta Heredia: Writing – review & editing, Software, Methodology, Investigation. Carlos Pérez de Arenaza García: Writing – review & editing, Software, Methodology, Investigation. Shuoyu Xu: Writing – review & editing, Software, Methodology, Investigation. Bingdou He: Writing – review & editing, Software, Methodology, Investigation. Xinping Cheng: Writing – review & editing, Software, Methodology, Investigation. Tao Wang: Writing – review & editing, Software, Methodology, Investigation. Noémie Moreau: Writing – review & editing, Software, Methodology, Investigation, Conceptualization. Katarzyna Bozek: Writing – review & editing, Software, Methodology, Investigation. Shubham Innani: Writing – review & editing, Software, Methodology, Investigation. Ujjwal Baid: Writing – review & editing, Software, Methodology, Investigation. Kaura Solomon Kefas: Writing – review & editing, Software, Methodology, Investigation. Bennett A. Landman: Writing – review & editing, Writing – original draft, Supervision, Resources, Project administration, Funding acquisition, Conceptualization. Yu Wang: Software, Resources, Investigation, Data curation. Shilin Zhao: Supervision, Resources, Project administration, Data curation. Mengmeng Yin: Validation, Supervision, Resources, Project administration, Data curation. Haichun Yang: Writing – review & editing, Writing – original draft, Supervision, Resources, Project administration, Investigation, Funding acquisition, Data curation, Conceptualization. Yuankai Huo: Writing – review & editing, Writing – original draft, Supervision, Resources, Project administration, Methodology, Investigation, Funding acquisition, Data curation, Conceptualization.

Data availability

KPIs 2024 Challenge data is publicly available on Synapse; code for the participating methods is at the GitHub repositories cited in the manuscript.

References

  1. Agarwal S, Sethi S, Dinda A, 2013. Basics of kidney biopsy: A nephrologist’s perspective. Indian J. Nephrol. 23 (4), 243–252. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Al Tuhaifi T, Zhong J, Yang H-C, Fogo AB, 2024. Effects of dipeptidyl peptidase-4 inhibitor and angiotensin-converting enzyme inhibitor on experimental diabetic kidney disease. Lab. Invest. 104 (2), 100305. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Altini N, Cascarano GD, Brunetti A, Marino F, Rocchetti MT, Matino S, Venere U, Rossini M, Pesce F, Gesualdo L, Bevilacqua V, 2020. Semantic segmentation framework for glomeruli detection and classification in kidney histological sections. Electronics 9 (3), 10.3390/electronics9030503, URL: https://www.mdpi.com/2079-9292/9/3/503. [DOI] [Google Scholar]
  4. Bandi P, Geessink O, Manson Q, Van Dijk M, Balkenhol M, Hermsen M, Bejnordi BE, Lee B, Paeng K, Zhong A, et al. , 2018. From detection of individual metastases to classification of lymph node status at the patient level: the camelyon17 challenge. IEEE Trans. Med. Imaging 38 (2), 550–560. [DOI] [PubMed] [Google Scholar]
  5. Bándi P, Geessink O, Manson Q, Van Dijk M, Balkenhol M, Hermsen M, Ehteshami Bejnordi B, Lee B, Paeng K, Zhong A, Li Q, Zanjani FG, Zinger S, Fukuta K, Komura D, Ovtcharov V, Cheng S, Zeng S, Thagaard J, Dahl AB, Lin H, Chen H, Jacobsson L, Hedlund M, Çetin M, Halıcı E, Jackson H, Chen R, Both F, Franke J, Küsters-Vandevelde H, Vreuls W, Bult P, van Ginneken B, van der Laak J, Litjens G, 2019. From detection of individual metastases to classification of lymph node status at the patient level: The CAMELYON17 challenge. IEEE Trans. Med. Imaging 38 (2), 550–560. 10.1109/TMI.2018.2867350. [DOI] [PubMed] [Google Scholar]
  6. Bankhead P, Loughrey MB, Fernández JA, Dombrowski Y, McArt DG, Dunne PD, McQuaid S, Gray RT, Murray LJ, Coleman HG, et al. , 2017. QuPath: Open source software for digital pathology image analysis. Sci. Rep. 7 (1), 1–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Barisoni L, Nast CC, Jennette JC, Hodgin JB, Herzenberg AM, Lemley KV, Conway CM, Kopp JB, Kretzler M, Lienczewski C, Avila-Casado C, Bagnasco S, Sethi S, Tomaszewski J, Gasim AH, Hewitt SM, 2013. Digital pathology evaluation in the multicenter Nephrotic Syndrome Study Network (NEPTUNE). Clin. J. Am. Soc. Nephrol. 8 (8), 1449–1459. 10.2215/CJN.08370812. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Barisoni L, Troost JP, Nast C, Bagnasco S, Avila-Casado C, Hodgin J, Palmer M, Rosenberg A, Gasim A, Liensziewski C, Merlino L, Chien H-P, Chang A, Meehan SM, Gaut J, Song P, Holzman L, Gibson D, Kretzler M, Gillespie BW, Hewitt SM, 2016. Reproducibility of the NEPTUNE descriptor-based scoring system on whole-slide images and histologic and ultrastructural digital images. Mod. Pathol. 29 (7), 671–684. 10.1038/modpathol.2016.58, URL: https://www.sciencedirect.com/science/article/pii/S0893395222021354. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Becker JU, Mayerich D, Padmanabhan M, Barratt J, Ernst A, Boor P, Cicalese PA, Mohan C, Nguyen HV, Roysam B, 2020. Artificial intelligence and machine learning in nephropathology. Kidney Int. 98 (1), 65–75. 10.1016/j.kint.2020.02.027, URL: https://www.sciencedirect.com/science/article/pii/S0085253820303422. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Bel T.d., Hermsen M, Litjens G, Laak J, 2018. Structure instance segmentation in renal tissue: a case study on tubular immune cell detection. In: Computational Pathology and Ophthalmic Medical Image Analysis. Springer, pp. 112–119. [Google Scholar]
  11. Bengtsson E, Danielsen H, Treanor D, Gurcan MN, MacAulay C, Molnár B, 2017. Computer-aided diagnostics in digital pathology. [DOI] [PubMed] [Google Scholar]
  12. Brixtel R, Bougleux S, Lézoray O, Caillot Y, Lemoine B, Fontaine M, Nebati D, Renouf A, 2022. Whole slide image quality in digital pathology: Review and perspectives. IEEE Access 10, 131005–131035. 10.1109/ACCESS.2022.3227437. [DOI] [Google Scholar]
  13. Bueno G, Fernandez-Carrobles MM, Gonzalez-Lopez L, Deniz O, 2020. Glomerulosclerosis identification in whole slide images using semantic segmentation. Comput. Methods Programs Biomed. 184, 105273. 10.1016/j.cmpb.2019.105273, URL: https://www.sciencedirect.com/science/article/pii/S0169260719311381. [DOI] [PubMed] [Google Scholar]
  14. Bulten W, Kartasalo K, Chen P-HC, Ström P, Pinckaers H, Nagpal K, Cai Y, Steiner DF, van Boven H, Vink R, Hulsbergen-van de Kaa C, van der Laak J, et al. , 2022b. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nature Med. 28 (1), 154–163. 10.1038/s41591-021-01620-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Bulten W, Kartasalo K, Chen P-HC, Ström P, Pinckaers H, Nagpal K, Cai Y, Steiner DF, Van Boven H, Vink R, et al. , 2022a. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nature Med. 28 (1), 154–163. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Cai P, Zhang X, Wang X, Du B, Lan L, Zhao Z, 2024. A robust deep learning method for WSI-level diseased glomeruli segmentation. In: International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis. Springer, pp. 192–201. [Google Scholar]
  17. Cap QH, 2024. An effective pipeline for whole-slide image glomerulus segmentation. arXiv preprint arXiv:2411.04782. [Google Scholar]
  18. Chen Y, Li W, Li Z, Sun R, Zhang T, Xiong Z, Wu F, 2024. Sam-glomeruli: Enhanced segment anything model for precise glomeruli segmentation. In: International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis. Springer, pp. 182–191. [Google Scholar]
  19. Cirillo L, Innocenti S, Becherucci F, 2024. Global epidemiology of kidney cancer. Nephrol. Dial. Transplant. 39 (6), 920–928. [DOI] [PubMed] [Google Scholar]
  20. Cui C, Deng R, Guo J, Liu Q, Yao T, Yang H, Huo Y, 2024. Enhancing physician flexibility: Prompt-guided multi-class pathological segmentation for diverse outcomes. In: IEEE-EMBS International Conference on Biomedical and Health Informatics. [Google Scholar]
  21. Deng R, Cui C, Liu Q, Yao T, Remedios LW, Bao S, Landman BA, Wheless LE, Coburn LA, Wilson KT, et al. , 2023a. Segment anything model (sam) for digital pathology: Assess zero-shot segmentation on whole slide imaging. arXiv preprint arXiv:2304.04155. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Deng R, Liu Q, Cui C, Yao T, Long J, Asad Z, Womick RM, Zhu Z, Fogo AB, Zhao S, et al. , 2023b. Omni-seg: A scale-aware dynamic network for renal pathological image segmentation. IEEE Trans. Biomed. Eng. 70 (9), 2636–2644. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Deng R, Liu Q, Cui C, Yao T, Xiong J, Bao S, Li H, Yin M, Wang Y, Zhao S, et al. , 2024a. HATs: Hierarchical adaptive taxonomy segmentation for panoramic pathology image analysis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, pp. 155–166. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Deng R, Liu Q, Cui C, Yao T, Yue J, Xiong J, Yu L, Wu Y, Yin M, Wang Y, et al. , 2024b. PrPSeg: Universal proposition learning for panoramic renal pathology segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11736–11746. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Ding H, Pan Z, Cen Q, Li Y, Chen S, 2020. Multi-scale fully convolutional network for gland segmentation using three-class classification. Neurocomputing 380, 150–161. [Google Scholar]
  26. Dosovitskiy A, 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. [Google Scholar]
  27. Ehteshami Bejnordi B, Litjens G, Timofeeva N, Otte-Höller I, Homeyer A, Karssemeijer N, van der Laak JA, 2016. Stain specific standardization of whole-slide histopathological images. IEEE Trans. Med. Imaging 35 (2), 404–415. 10.1109/TMI.2015.2476509. [DOI] [PubMed] [Google Scholar]
  28. Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens G, van der Laak JAWM, the CAMELYON16 Consortium, 2017. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA 318 (22), 2199–2210. 10.1001/jama.2017.14585, arXiv:https://jamanetwork.com/journals/jama/articlepdf/2665774/jama_ehteshami_bejnordi_2017_oi_170113.pdf. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Gallego J, Swiderska-Chadaj Z, Markiewicz T, Yamashita M, Gabaldon MA, Gertych A, 2021. A U-Net based framework to quantify glomerulosclerosis in digitized PAS and H&E stained human tissues. Comput. Med. Imaging Graph. 89, 101865. 10.1016/j.compmedimag.2021.101865, URL: https://www.sciencedirect.com/science/article/pii/S0895611121000136. [DOI] [PubMed] [Google Scholar]
  30. Ginley B, Lutnick B, Jen K-Y, Fogo AB, Jain S, Rosenberg A, Walavalkar V, Wilding G, Tomaszewski JE, Yacoub R, et al. , 2019. Computational segmentation and classification of diabetic glomerulosclerosis. J. Am. Soc. Nephrol. 30 (10), 1953–1967. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Ginley B, Tomaszewski JE, Sarder P, 2017. Automatic computational labeling of glomerular textural boundaries. In: Medical Imaging 2017: Digital Pathology. Vol. 10140, SPIE, pp. 100–105. [Google Scholar]
  32. Gomes J, Kong J, Kurc T, Melo AC, Ferreira R, Saltz JH, Teodoro G, 2021. Building robust pathology image analyses with uncertainty quantification. Comput. Methods Programs Biomed. 208, 106291. [DOI] [PubMed] [Google Scholar]
  33. Guo M-H, Lu C-Z, Hou Q, Liu Z, Cheng M-M, Hu S-M, 2022. Segnext: Rethinking convolutional attention design for semantic segmentation. Adv. Neural Inf. Process. Syst. 35, 1140–1156. [Google Scholar]
  34. He K, Zhang X, Ren S, Sun J, 2016. Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770–778. [Google Scholar]
  35. Hogan JJ, Mocanu M, Berns JS, 2016. The native kidney biopsy: update and evidence for best practice. Clin. J. Am. Soc. Nephrol. 11 (2), 354–362. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Hou X, Wu EZ, Dong G, Yuan W, Li Y, Wang S, 2024. AC-UNet: a self-adaptive cropping approach for kidney pathology image segmentation. In: International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis. Springer, pp. 173–181. [Google Scholar]
  37. Hu EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, Wang L, Chen W, 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv: 2106.09685. [Google Scholar]
  38. Huang H, Lin L, Tong R, Hu H, Zhang Q, Iwamoto Y, Han X, Chen Y-W, Wu J, 2020. Unet 3+: A full-scale connected unet for medical image segmentation. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing. ICASSP, IEEE, pp. 1055–1059. [Google Scholar]
  39. Iglesias P, Carrero JJ, Díez JJ, 2012. Gonadal dysfunction in men with chronic kidney disease: clinical features, prognostic implications and therapeutic options. J. Nephrol. 25 (1), 31. [DOI] [PubMed] [Google Scholar]
  40. Isensee F, Jaeger PF, Kohl SA, Petersen J, Maier-Hein KH, 2021. NnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18 (2), 203–211. [DOI] [PubMed] [Google Scholar]
  41. Jain MN, Al-Heejawi SMA, Azzi JR, Amal S, 2025. Digital pathology and ensemble deep learning for kidney cancer diagnosis: Dartmouth kidney cancer histology dataset. Appl. Biosci. 4 (1), 8. [Google Scholar]
  42. Jain Y, Godwin LL, Ju Y, Sood N, Quardokus EM, Bueckle A, Longacre T, Horning A, Lin Y, Esplin ED, Hickey JW, Snyder MP, Patterson NH, Spraggins JM, Börner K, 2023. Segmentation of human functional tissue units in support of a Human Reference Atlas. Commun. Biol. 6 (1), 717. 10.1038/s42003-023-04848-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Jha A, Yang H, Deng R, Kapp ME, Fogo AB, Huo Y, 2021. Instance segmentation for whole slide imaging: end-to-end or detect-then-segment. J. Med. Imaging 8 (1), 014001–014001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Jiang L, Chen W, Dong B, Mei K, Zhu C, Liu J, Cai M, Yan Y, Wang G, Zuo L, Shi H, 2021. A deep learning-based approach for glomeruli instance segmentation from multistained renal biopsy pathologic images. Am. J. Pathol. 191 (8), 1431–1441. 10.1016/j.ajpath.2021.05.004, URL: https://www.sciencedirect.com/science/article/pii/S0002944021002030. [DOI] [PubMed] [Google Scholar]
  45. Jiménez-Heffernan J, Bajo MA, Perna C, del Peso G, Larrubia JR, Gamallo C, Sánchez-Tomero J, López-Cabrera M, Selgas R, 2006. Mast cell quantification in normal peritoneum and during peritoneal dialysis treatment. Arch. Pathol. Lab. Med. 130 (8), 1188–1192. [DOI] [PubMed] [Google Scholar]
  46. Kellum JA, 2008. Acute kidney injury. Crit. Care Med. 36 (4), S141–S145. [DOI] [PubMed] [Google Scholar]
  47. Kidney Precision Medicine Project, 2026. Kidney precision medicine project. The results here are in whole or part based upon data generated by the KPMP. https://www.kpmp.org. (Accessed 02 September 2026).
  48. Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, Xiao T, Whitehead S, Berg AC, Lo W-Y, et al. , 2023. Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026. [Google Scholar]
  49. Kumar A, Das DK, Deotale G, Dalimkar V, Thomas T, Singhal N, 2024. Ensembled SegNeXt based glomeruli segmentation. In: International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis. Springer, pp. 202–209. [Google Scholar]
  50. Kumar N, Verma R, Anand D, Zhou Y, Onder OF, Tsougenis E, Chen H, Heng P-A, Li J, Hu Z, Wang Y, Koohbanani NA, Jahanifar M, Tajeddin NZ, Gooya A, Rajpoot N, Ren X, Zhou S, Wang Q, Shen D, Yang C-K, Weng C-H, Yu W-H, Yeh C-Y, Yang S, Xu S, Yeung PH, Sun P, Mahbod A, Schaefer G, Ellinger I, Ecker R, Smedby O, Wang C, Chidester B, Ton T-V, Tran M-T, Ma J, Do MN, Graham S, Vu QD, Kwak JT, Gunda A, Chunduri R, Hu C, Zhou X, Lotfi D, Safdari R, Kascenas A, O’Neil A, Eschweiler D, Stegmaier J, Cui Y, Yin B, Chen K, Tian X, Gruening P, Barth E, Arbel E, Remer I, Ben-Dor A, Sirazitdinova E, Kohl M, Braunewell S, Li Y, Xie X, Shen L, Ma J, Baksi KD, Khan MA, Choo J, Colomer A, Naranjo V, Pei L, Iftekharuddin KM, Roy K, Bhattacharjee D, Pedraza A, Bueno MG, Devanathan S, Radhakrishnan S, Koduganty P, Wu Z, Cai G, Liu X, Wang Y, Sethi A, 2020. A multi-organ nucleus segmentation challenge. IEEE Trans. Med. Imaging 39 (5), 1380–1391. 10.1109/TMI.2019.2947628. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Kumar N, Verma R, Anand D, Zhou Y, Onder OF, Tsougenis E, Chen H, Heng P-A, Li J, Hu Z, et al. , 2019. A multi-organ nucleus segmentation challenge. IEEE Trans. Med. Imaging 39 (5), 1380–1391. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Kumar N, Verma R, Sharma S, Bhargava S, Vahadane A, Sethi A, 2017. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Trans. Med. Imaging 36 (7), 1550–1560. [DOI] [PubMed] [Google Scholar]
  53. Lim BJ, Yang H-C, Fogo AB, 2014. Animal models of regression/progression of kidney disease. Drug Discov. Today: Dis. Model. 11, 45–51. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Lim BJ, Yang JW, Zou J, Zhong J, Matsusaka T, Pastan I, Zhang M-Z, Harris RC, Yang H-C, Fogo AB, 2017. Tubulointerstitial fibrosis can sensitize the kidney to subsequent glomerular injury. Kidney Int. 92 (6), 1395–1403. [DOI] [PMC free article] [PubMed] [Google Scholar]
  55. Lin T-Y, Dollár P, Girshick R, He K, Hariharan B, Belongie S, 2017. Feature pyramid networks for object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2117–2125. [Google Scholar]
  56. Litjens G, Bandi P, Ehteshami Bejnordi B, Geessink O, Balkenhol M, Bult P, Halilovic A, Hermsen M, van de Loo R, Vogels R, Manson QF, Stathonikos N, Baidoshvili A, van Diest P, Wauters C, van Dijk M, van der Laak J, 2018. 1399 H&E-stained sentinel lymph node sections of breast cancer patients: the CAMELYON dataset. GigaScience 7 (6), giy065. 10.1093/gigascience/giy065, arXiv:https://academic.oup.com/gigascience/article-pdf/7/6/giy065/60709038/gigascience_7_6_giy065.pdf. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Liu Y, 2024. A hybrid CNN-TransXNet approach for advanced glomerular segmentation in renal histology imaging. Int. J. Comput. Intell. Syst. 17 (1), 126. 10.1007/s44196-024-00523-7. [DOI] [Google Scholar]
  58. Liu Z, Mao H, Wu C-Y, Feichtenhofer C, Darrell T, Xie S, 2022. A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976–11986. [Google Scholar]
  59. Marti-Aguado D, Rodríguez-Ortega A, Mestre-Alagarda C, Bauza M, Valero-Pérez E, Alfaro-Cervello C, Benlloch S, Pérez-Rojas J, Ferrández A, Alemany-Monraval P, et al. , 2021. Digital pathology: accurate technique for quantitative assessment of histological features in metabolic-associated fatty liver disease. Aliment. Pharmacol. Ther. 53 (1), 160–171. [DOI] [PubMed] [Google Scholar]
  60. Mounier-Vehier C, Lions C, Devos P, Jaboureck O, Willoteaux S, Carre A, Beregi J-P, 2002. Cortical thickness: an early morphological marker of atherosclerotic renal disease. Kidney Int. 61 (2), 591–598. [DOI] [PubMed] [Google Scholar]
  61. Ren J, Sadimin E, Foran DJ, Qi X, 2017. Computer aided analysis of prostate histopathology images to support a refined Gleason grading system. In: Medical Imaging 2017: Image Processing. Vol. 10133, International Society for Optics and Photonics, 101331V. [DOI] [PMC free article] [PubMed] [Google Scholar]
  62. Ronneberger O, Fischer P, Brox T, 2015. U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer, pp. 234–241. [Google Scholar]
  63. Sánchez M, Sánchez H, Arenaza C.P.d., Ribalta D, Arrarte N, Cámara O, Galdran A, 2024. Glomeruli segmentation in whole-slide images: Is better local performance always better? In: International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis. Springer, pp. 210–217. [Google Scholar]
  64. Singh Samant S, Chauhan A, DN J, Singh V, 2023. Glomerulus detection using segmentation neural networks. J. Digit. Imaging 36, 1633–1642. 10.1007/s10278-022-00764-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  65. Tang Y, He Y, Nath V, Guo P, Deng R, Yao T, Liu Q, Cui C, Yin M, Xu Z, et al. , 2024. Holohisto: End-to-end gigapixel wsi segmentation with 4k resolution sequential tokenization. arXiv preprint arXiv:2407.03307. [Google Scholar]
  66. Wiesenfarth M, Reinke A, Landman BA, Eisenmann M, Saiz LA, Cardoso MJ, Maier-Hein L, Kopp-Schneider A, 2021. Methods and open-source toolkit for analyzing and visualizing challenge results. Sci. Rep. 11 (1), 2369. [DOI] [PMC free article] [PubMed] [Google Scholar]
  67. Wu B, Moeckel G, 2023. Application of digital pathology and machine learning in the liver, kidney and lung diseases. J. Pathol. Inform. 14, 100184. 10.1016/j.jpi.2022.100184, URL: https://www.sciencedirect.com/science/article/pii/S2153353922007842. [DOI] [PMC free article] [PubMed] [Google Scholar]
  68. Xiao T, Liu Y, Zhou B, Jiang Y, Sun J, 2018. Unified perceptual parsing for scene understanding. In: Proceedings of the European Conference on Computer Vision. ECCV, pp. 418–434. [Google Scholar]
  69. Xie E, Wang W, Yu Z, Anandkumar A, Alvarez JM, Luo P, 2021. SegFormer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inf. Process. Syst. 34, 12077–12090. [Google Scholar]
  70. Yao T, Lu Y, Long J, Jha A, Zhu Z, Asad Z, Yang H, Fogo AB, Huo Y, 2022. Glo-in-one: Holistic glomerular detection, segmentation, and lesion characterization with large-scale web image mining. arXiv:2206.00123, URL: https://arxiv.org/abs/2206.00123. [DOI] [PMC free article] [PubMed] [Google Scholar]
  71. Yin Y, Tang Z, Weng H, 2024. Application of visual transformer in renal image analysis. BioMed Eng. OnLine 23, 27. 10.1186/s12938-024-01209-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  72. Yu L, Yin M, Deng R, Liu Q, Yao T, Cui C, Guo J, Wang Y, Wang Y, Zhao S, Yang H, Huo Y, 2024. Glo-in-one-v2: Holistic identification of glomerular cells, tissues, and lesions in human and mouse histopathology. arXiv:2411.16961, URL: https://arxiv.org/abs/2411.16961. [DOI] [PMC free article] [PubMed] [Google Scholar]
  73. Yue J, Yao T, Deng R, Liu Q, Xiong J, Guo J, Yang H, Huo Y, 2024a. Weighted circle fusion: Ensembling circle representation from different object detection results. arXiv:2406.19540, URL: https://arxiv.org/abs/2406.19540. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Yue J, Yao T, Deng R, Lu S, Guo J, Liu Q, Yin M, Xiong J, Yang H, Huo Y, 2024b. GloFinder: AI-empowered QuPath plugin for WSI-level glomerular detection, visualization, and curation. arXiv:2411.18795, URL: https://arxiv.org/abs/2411.18795. [DOI] [PMC free article] [PubMed] [Google Scholar]
  75. Yun S, Han D, Oh SJ, Chun S, Choe J, Yoo Y, 2019. Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023–6032. [Google Scholar]
  76. Zeng C, Nan Y, Xu F, Lei Q, Li F, Chen T, Liang S, Hou X, Lv B, Liang D, et al. , 2020. Identification of glomerular lesions and intrinsic glomerular cell types in kidney diseases via deep learning. J. Pathol. 252 (1), 53–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
  77. Zhang H, 2017. Mixup: Beyond empirical risk minimization. arXiv preprint arXiv: 1710.09412. [Google Scholar]
  78. Zheng S, Lu J, Zhao H, Zhu X, Luo Z, Wang Y, Fu Y, Feng J, Xiang T, Torr PH, et al. , 2021. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6881–6890. [Google Scholar]
  79. Zhou Z, Rahman Siddiquee MM, Tajbakhsh N, Liang J, 2018. Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. Springer, pp. 3–11. [DOI] [PMC free article] [PubMed] [Google Scholar]
  80. Zhu J, Yin M, Deng R, Long Y, Wang Y, Wang Y, Zhao S, Yang H, Huo Y, 2025. Cross-species data integration for enhanced layer segmentation in kidney pathology. In: Medical Imaging 2025: Digital and Computational Pathology. Vol. 13413, SPIE, pp. 49–56. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

supplement 2
supplement 1

Data Availability Statement

KPIs 2024 Challenge data is publicly available on Synapse; code for the participating methods is at the GitHub repositories cited in the manuscript.

RESOURCES