A two-stage foundation model for bladder tumor segmentation: An international multi-site study

Abstract

Background and objective:

Accurate bladder tumor segmentation is crucial for muscle invasion assessment, neoadjuvant therapy response evaluation, and bladder preservation decision-making. Manual delineation is time-consuming and observer-dependent, highlighting an unmet clinical need for accurate and automated segmentation, particularly in large multi-site studies. This study aims to (i) develop an automated MRI-based foundation model for bladder tumor segmentation (BLA-T-Seg) using a two-stage framework, in which segmentation is first constrained to the bladder wall, including wall-contiguous tumors and then refined in a second stage focused specifically on the tumor, and (ii) rigorously assess generalizability via multi-site, independent external validation across diverse cohorts.

Methods:

This retrospective study included T2-weighted MRI scans from 236 patients with bladder cancer across two international sites. Experienced radiologists manually annotated bladder and tumor regions of interest on T2-weighted images, which served as the ground truth. BLA-T-Seg was developed by adapting the Segment Anything Model (SAM), fine-tuned to segment the bladder in stage 1 and the tumor in stage 2. The primary performance metric was the Dice similarity coefficient (DSC), evaluated across five studies (S1–S3: internal; S4–S5: external) and benchmarked against 10 state-of-the-art (SOTA) segmentation architectures.

Results:

For stage 1, BLA-T-Seg achieved good-to-excellent performance, with mean DSC values ranging from 0.85 to 0.93. For stage 2, BLA-T-Seg demonstrated moderate-to-good performance, with mean DSC values of 0.74–0.81. Cross-site validation showed minimal performance variation (≤0.01–0.04), confirmed by radar plot analyses. Compared with SOTA models on subset S1, BLA-T-Seg achieved the highest DSC of 0.84, outperforming the next best-performing model, Swin Transformer (DSC: 0.56).

Conclusion:

BLA-T-Seg enables accurate and generalizable bladder wall and tumor segmentation and outperforms SOTA alternatives.

Introduction

Artificial intelligence (AI)-driven image segmentation is transforming oncologic imaging by enabling accurate, automated, and reproducible delineation of tumors and organs at risk, thereby facilitating personalized treatment strategies [1–3]. Deep convolutional neural networks (CNN) have demonstrated a significant impact on image segmentation tasks by learning hierarchical spatial features directly from the data [4–6], while transformer-based architectures and self-configuring frameworks, have improved the segmentation performance and adaptability across heterogeneous medical datasets [7].

Recently, foundation models have reshaped the field of image segmentation by introducing large-scale pretraining paradigms designed to generalize across diverse downstream tasks [8–10]. These models leverage extensive multi-organ and multi-modal datasets to learn transferable visual representations that can be adapted to organ-specific applications. Ma et al. developed the MedSAM model, trained on approximately 1.5 million image-mask pairs across 10 imaging modalities and more than 30 cancer types, achieving broad segmentation capability across datasets [11]. Similarly, Dong et al. introduced a vision foundation model trained using more than 6 million slices from over 110,000 MRI volumes across 18 body locations [9]. Collectively, these studies established the feasibility of adapting foundation models for medical image segmentation at scale, showing that large pretrained models can provide transferable representations across diverse organs, modalities, and segmentation targets [9,11]. MedSAM, for example, was explicitly developed as a universal medical image segmentation model to address the limited generalizability of conventional task-specific approaches, and was evaluated across multiple internal and external tasks [11]. Similarly, large MRI foundation models have demonstrated the value of pretraining on heterogeneous MRI volumes for learning broadly applicable representations across anatomical regions [9]. However, these studies were primarily designed to evaluate broad segmentation generalizability rather than disease-specific oncologic segmentation workflows. In clinically focused segmentation tasks, performance depends not only on tumor localization but also on anatomical context, boundary ambiguity, lesion morphology, and inference strategies tailored to the target disease. Therefore, organ-specific adaptation and validation remain necessary to determine whether foundation-model approaches can support reliable segmentation in clinically meaningful MRI-based oncology applications. The robust generalization across institutions, scanner vendors, and acquisition protocols remains a major barrier to clinical translation of segmentation models.

Despite promising advances, the application of foundation models to MRI-based bladder tumor segmentation has not yet been investigated. Bladder MRI plays an increasingly central role in muscle invasion assessment, neoadjuvant treatment response evaluation, and bladder preservation decision-making. Bladder tumors exhibit marked morphological heterogeneity, including papillary, sessile, broad-based, and multifocal growth patterns, often with irregular intraluminal projections [12–15]. The diversity in tumor shape, size, and strong intensity inhomogeneity in urine caused by motion artifacts, weak boundaries, and complex background intensity distribution make tumor segmentations challenging on MRI [16]. The manual segmentation remains labor-intensive and prone to inter-observer variability because tumors may arise from and remain contiguous with the bladder wall, prone to boundary ambiguity. These combined anatomical and procedural challenges underscore the need for automated segmentation systems that can deliver consistent and reproducible performance independent of reader expertise. Further, rigorous multi-site validation in geographically distinct populations is needed to establish generalizability and translational reliability for MRI-based bladder tumor segmentation.

To address these unmet clinical and technical challenges, this study has two primary aims (i) develop a fully automated MRI-based foundation model for bladder tumor segmentation (BLA-T-Seg) using a two-stage approach in which segmentation is first constrained to the bladder wall, including tumor boundaries contiguous with the wall, and then refined in a second stage focused specifically on the tumor to promote structurally consistent and reproducible segmentation, and (ii) rigorously evaluate the generalizability and robustness of the proposed model through a multi-site study design, including independent external validation across diverse patient cohorts.

Material and methods

Study sample

This retrospective multi-site study was approved by the institutional review board at Memorial Sloan Kettering, a cancer center in the United States, and was compliant with the Health Insurance Portability and Accountability Act. The need for informed consent was waived. To develop BLA-T-Seg, an internal dataset was used, comprising axial T2-weighted images from MRI examinations performed between July 2015 and April 2024 [17]. Inclusion criteria were (i) muscle invasive bladder cancer (MIBC) patients who had pathologic and/or radiologic confirmed clinical stage II-III (clinical stage T2–4a N1 M0) disease, patients with cT4a disease invading the prostatic stroma with no cystoscopic confirmation of muscle invasion and (ii) histopathologically confirmed non-(N)MIBC (clinical stage Ta, T1, or CIS) based on Transurethral Resection of Bladder Tumor (TURBT). All patients who could be scanned with MRI protocol were eligible for the study. Exclusion criteria were (i) patients with a known contraindication to MRI and (ii) patients with any component of small cell/neuroendocrine bladder cancer and pure variant histology. To evaluate the generalizability and translational reliability for BLA-T-Seg, an external dataset was included from IRCCS San Raffaele Hospital, Comprehensive Cancer Center, in Italy [18]. The external dataset comprised T2-weighted images from MRI examinations performed between February 2017 and December 2019. Inclusion criteria and Exclusion criteria were similar to those for MIBC mentioned above. Subject overlap with previous treatment response studies is detailed in supplementary material P1.

MRI acquisition parameters

All MRI acquisitions at the internal site were performed using a 1.5 T or 3 T GE HealthCare scanner, while MRI acquisitions at the external site were performed using a 1.5 T Philips scanner. Axial T2-weighted imaging sequences were performed with the following parameters: at the internal site, repetition time (TR)/echo time (TE) = 3000–4500/90–120 ms, matrix size = 320 × 256, field of view (FOV) = 200–240 mm, slice thickness = 3–5 mm, and number of slices = 30–35; and at the external site, TR/TE = 3500–4000/90–110 ms, matrix size = 320 × 256, FOV = 200–220 mm, slice thickness = 3–4 mm, and number of slices = 40–45.

Manual segmentations to serve as reference inputs for BLA-T-Seg development

T2-weighted images were retrieved from the institutional Picture Archiving and Communication System (PACS) in DICOM format, de-identified and subsequently imported into XNAT for manual segmentation of tumor and bladder regions of interest (ROIs).

Manual tumor segmentations were performed by four genitourinary radiologists with expertise in bladder MRI, including three junior radiologists (GB, SC, JN) > 5 years of experience) and one senior radiologist (OA, > 20 years of experience). Each T2-weighted image was independently segmented by a junior-senior radiologist pair. Disagreement and discrepancies between readers were resolved through consensus review, and final segmentations were confirmed by the paired radiologists to ensure consistency and quality.

Separately, to anatomically constrain the BLA-T-Seg pipeline and enable reliable bladder localization prior to downstream tumor auto-segmentation, following manual tumor segmentations, manual bladder segmentations were performed on the same T2-weighted images by two imaging scientists (*MA and RP, > 2 years of experience) to include tumor boundaries contiguous with the bladder wall. On T2-weighted images, urine appears markedly hyperintense relative to the hypointense muscular wall, providing a naturally high-contrast boundary that facilitates consistent and accurate organ localization. Thus, the imaging scientists leveraged this inherent contrast to delineate the bladder ROIs, covering most of the bladder (± 3–5 slices superior and inferior in the z-direction).

Two-stage BLA-T-Seg development pipeline

Overall strategy and data partitioning

A two-stage training strategy for BLA-T-Seg was applied, where stage 1 involves bladder wall (including contiguous tumor) segmentation, and stage 2 involves tumor segmentation. Fig. 1 provides an overview of the BLA-T-Seg development pipeline. A strict data partitioning scheme was implemented to ensure full independence between training stages.

Fig. 1.

Workflow of the BLA-T-Seg framework, encompassing data preparation, preprocessing, and automated segmentation. (A) Data preparation: MRI data of patients with bladder cancer were collected from the United States and Italy. Expert radiologists manually annotated the bladder and tumor regions of interest on T2-weighted images to generate reference masks. (B) Data preprocessing: Three-dimensional NIfTI images, including whole T2-weighted images and corresponding bladder and tumor references masks, underwent standardized preprocessing, including N4 bias field correction and intensity normalization following Imaging Biomarker Standardization Initiative guidelines. (C) Automated segmentation: A two-stage deep learning segmentation pipeline was employed. First, an AI model was trained to automatically segment the bladder. Post-processing operations were applied to refine the segmentation and extract the bladder region of interest, which was subsequently used to train a second AI model for precise bladder tumor segmentation. BLA-T-Seg performance was evaluated on internal and external data subsets.

Stage 1 – bladder wall, including tumor boundaries when contiguous with the wall segmentation:

S1 was split into a training set (80%) and a validation set (20%). S2 and S3 served as internal test sets, while S4 and S5 served as external test sets.

Stage 2 – tumor segmentation:

The Stage 1 validation set (remaining 20% of S1), combined with S2 and S3, was repartitioned into a training set (80%) and a validation set (20%). S4 and S5 were again used as external test sets.

Data preprocessing

Given that MRI signal intensities are expressed in arbitrary units and vary across scanners, vendors, and acquisition protocols, T2-weighted images, converted from raw DICOM format to NIfTI format through a dedicated interface [19], along with manual tumor and bladder annotation masks, underwent standardized preprocessing, using the previously described AI-BLADE toolbox [20]. First, N4 bias field correction was performed to reduce low-frequency intensity inhomogeneity caused by magnetic field non-uniformities, which can adversely affect quantitative analysis and model performance. Following bias correction, image intensities were normalized using robust z-score standardization, defined as $\left(0.6745\mathrm{*}\right(\mathrm{x}\_\mathrm{i}-\tilde{\mathrm{x}})/\mathrm{M}\mathrm{A}\mathrm{D}$, where $\mathrm{x}\_\mathrm{i}$ is voxel intensity, $\tilde{\mathrm{x}}$, mean absolute deviation of voxel intensities and MAD is the median absolute deviation computed over the image volume. This robust normalization approach reduces sensitivity to outliers and intensity extremes, enabling more stable feature learning across heterogeneous MRI datasets.

Segment anything model (SAM) fine-tuning

The Segment Anything Model (SAM) [8] was fine-tuned for BLA-T-Seg, using the Vision Transformer Large (SAM-ViT-L) backbone through the HuggingFace implementation [21]. Fig. 2 provides a schematic of the fine-tuning pipeline. Since SAM was originally designed for natural images, a series of preprocessing and integration steps were required to make the framework compatible with our cancer imaging data prior to fine-tuning. As SAM typically relies on prompts (points, boxes, or rough masks), we designed a custom PyTorch collate function to tailor the data to our use case. All MRI slices and their corresponding masks were processed using the SamProcessor, which standardizes inputs to a fixed 256 × 256 resolution. and further formats each into the structure required by SAM, generating pixel tensors alongside empty prompt embeddings. Then, for each training sample, a custom PyTorch collate function loaded the T2-weighted images and their corresponding tumor masks, applied all SamProcessor transformations, and assembled the processed inputs into batches. SAM was intentionally trained in a prompt-free mode, forcing the model to rely solely on image context to generate tumor masks.

Fig. 2.

Overview of the Segment Anything Model (SAM) fine-tuning pipeline for two-stage segmentation: Stage 1 – bladder wall segmentation, including tumor boundaries when contiguous with the wall; and Stage 2 – tumor segmentation.

After preprocessing, the dataset was partitioned into training and testing subsets according to a predefined split ratio. PyTorch DataLoaders were configured with a batch size of 1 to accommodate the substantial memory requirements of the SAM-ViT-L architecture and were executed with multiple CPU workers to improve data loading efficiency. Of note, as SAM requires three-channel inputs, the original single-channel images were appropriately mapped to a three-channel format ensuring full compatibility with the network.

A full fine-tuning regime was adopted, in which all components of the SAM architecture, including the vision encoder, prompt encoder, and mask decoder, were optimized jointly. Training was performed using the Adam optimizer with a learning rate of 1 × 10⁻⁵ and no weight decay, in conjunction with a combined Dice and cross-entropy loss function. This loss formulation balances region-level overlap through the Dice term with voxel-wise classification accuracy via the cross-entropy term, making it well suited for medical image segmentation tasks involving small or irregular tumor boundaries. Sigmoid activation and squared predictions were enabled to stabilize the Dice component, and mean reduction was applied to ensure consistent gradient behavior.

Data postprocessing

To enhance the anatomical consistency of both the bladder wall segmentation, including tumor boundaries when contiguous with the wall, and the subsequent tumor segmentation, we applied a dedicated post-processing pipeline to all predicted masks. This procedure was intended to remove small noisy regions, suppress spurious detections, and preserve only the most anatomically relevant connected components.

Thresholding:

Predicted probability maps were first converted to binary masks. Let $I(x,y)$ denote the grayscale or probabilistic value at pixel ($x,y$). We defined the binary mask $B(x,y)$ using a fixed threshold:

$$B(x,y)=\left\{\begin{array}{ll}255,& \mathit{if}I(x,y)\ge 128,\\ & 0,\mathit{other}\mathit{wise}\end{array}\right.$$

where 255 corresponds to foreground and 0 to background. This step enforces a hard separation between object and background while preserving high-confidence regions.

Morphological Filtering:

To remove isolated noise and small gaps, we applied a morphological opening operation to the binary mask. Opening consists of an erosion followed by a dilation using a $3x3$ square structuring element

$$k=\left[\begin{array}{ll}1& 1\\ 1& 1\\ 1& 1\end{array}\right]$$

Given a binary mask $B$, the opened mask $O\left(B\right)$ is defined as

$$O\left(B\right)=D\left(E\right(B,K),K),$$

where $E(.,K)$ and $D(.,K)$ denote erosion and dilation with kernel $K$ respectively. This operation removes small, isolated foreground components and smooths the boundary of larger structures, which is particularly important for suppressing speckle-like noise around the bladder wall.

Connected Component Analysis:

Next, we performed connected component analysis on the opened mask $O\left(B\right)$ to separate individual foreground regions. Let $\left\{R_i\right\}$ denote the set of all connected components (white regions) and $A_i$ their corresponding areas in pixels. We computed the empirical distribution of ($A_i$) and determined the 75th percentile $P_{75}\left(A\right)$.Components were then filtered by area such that only the largest 25% were retained:

$$A_t=P_{75}(A),R_i^{\mathrm{*}}=R_i\mid A_i\ge A_t$$

where $A_t$ is the area threshold and $(R_i^{\mathrm{*}})$ is the set of retained components. This percentile-based criterion suppresses small, anatomically implausible regions while preserving the bladder or tumor structures.

Mask Reconstruction:

Finally, a refined binary mask $M_{\mathit{final}}$ was reconstructed by combining all retained components:

$$M_{\mathit{final}}=\left\{\begin{array}{ll}255& ,\mathit{if}(x,y)\in U_i{R}_i^{\mathrm{*}}\\ & 0,\mathit{Otherwise}\end{array}\right.$$

The resulting masks contain only the most significant connected regions and exhibit smoother, more contiguous boundaries.

All post-processed ROI masks were saved to the output directory with the original filenames preserved. This post-processing procedure was applied uniformly across all segmentation outputs. Within the two-stage pipeline, bladder masks encompassing the bladder wall and any contiguous tumor were first refined using the above procedure, after which the cropped ROIs facilitated targeted and reproducible bladder tumor delineation. Overall, this step played a central role in stabilizing the hierarchical segmentation framework; first to the bladder wall, including tumor boundaries when contiguous with the wall, and then to the tumor itself, while reducing false-positive findings arising from small, noisy regions.

All experiments were implemented using Python with PyTorch, MONAI, and HuggingFace Transformers libraries. Training and inference were conducted on NVIDIA GPU-based systems to ensure computational reproducibility.

Statistical analysis

BLA-T-Seg was evaluated using multiple quantitative metrics (Dice similarity coefficient [DSC], intersection over union [IoU], precision, recall, F1-score, and accuracy) to provide a comprehensive assessment of spatial overlap, boundary agreement, and voxel-wise classification accuracy [22]. The primary metric, the DSC, measures volumetric overlap between the predicted segmentation and the ground truth reference.

To facilitate consistent interpretation across experiments, DSC values were categorized based on predefined thresholds as follows: values < 0.60 were considered indicative of poor agreement, 0.60–0.69 as fair agreement, 0.70–0.79 as moderate agreement, 0.80–0.89 as good agreement, and 0.90–1.00 as excellent or near-perfect overlap. These categories were adopted to standardize reporting and ensure consistency in the description of segmentation performance across the Results section. Accordingly, qualitative descriptors (e.g., “good” or “excellent”) were used in alignment with these predefined DSC ranges.

In addition, Intersection over Union (IoU) was computed to further assess region-level overlap robustness. Voxel-wise classification performance was evaluated using accuracy, precision, sensitivity (recall), and specificity. Precision quantified the proportion of correctly predicted positive voxels among all predicted positives, while sensitivity measured the model’s ability to correctly identify true tumor voxels. Specificity assessed correct identification of non-tumor regions, and overall accuracy reflected the proportion of correctly classified voxels across the entire image volume.

All metrics were calculated separately for segmentation to the bladder wall (including tumor boundaries when contiguous with the wall) followed by tumor segmentation tasks and summarized as mean ± standard deviation across internal held-out test sets and the independent external cohort. Comparative analysis between internal and external cohorts was performed to evaluate cross-site robustness and generalizability of the proposed BLA-T-Seg framework.

All statistical analyses were performed using Python (version 3.10; Python Software Foundation, Wilmington, Delaware) with standard scientific libraries (NumPy, SciPy, scikit-learn, and PyTorch).

To benchmark BLA-T-Seg’s segmentation performance, the segmentation performance of ten widely used state-of-the-art deep learning architectures was also evaluated, including CNN-based architectures (Attention U-Net [23], DeepLabv3 + [24], U-Net [25], U-Net-ASPP [26], SegformerForSemanticSegmentation [26], UNet3 + [27], and UNETR [7]), transformer-based approaches (Swin Transformer [28] and Swin UNETR [29], as well as a pretrained foundation model (MRI-CORE [9]).

Results

Patient characteristics

A total of 236 patients were included and stratified into five subsets. The internal cohorts consisted of S1 (n = 106), comprising pretreatment MRI from patients with MIBC treated with cisplatin-based neoadjuvant chemotherapy (NAC); S2 (n = 27), post-NAC MRI from MIBC patients; and S3 (n = 35), pretreatment MRI from patients with NMIBC. The external cohorts consisted of S4 (n = 68) and S5 (n = 68), which included pre- and post-treatment MRI from MIBC patients treated with immunotherapy. Table 1 summarizes the study sample characteristics. Median age (interquartile range [IQR]) was 66 years [59–70] in S1, 64 [57–70] in S2, 67 years [60–73] in S3, and 68 years [62–73] in S4–5.

Table 1.

Patient characteristics across five study subsets.

Characteristic	S1 (n = 106)	S2 (n = 27)	S3* (n = 35)	S4–5 (n = 68)
Age (y) (Median (IQR))	66(59–70	64 [57–70]	67 [60–73]	68 [62, 73]
Sex (Male)	69 (65%)	25 (93%)	28 (80%)	56 (82%)
Smoking
Former	60 (57%)	17 (63%)	12 (34%)	27 (40%)
Never	34 (32%)	6 (22%)	9 (26%)	22 (32%)
Current	11 (10%)	3 (11%)	6 (17%)	19 (28%)
Unknown	1 (1%)	1 (4%)	6 (17%)
Histology
Non-predominant VH	40 (38%)	8 (30%)	N/A	6 (9%)
Predominant VH	2 (2.0%)	1 (3%)	N/A	10 (15%)
Pure UC	64 (60%)	18 (67%)	33 (94%)	52 (76%)

Note: The five study subsets are: S1-S3 comprising internal study samples; S4-S5 comprising external study samples. Abbreviations: UC = urothelial carcinoma;

VH = variant histology

^*Information for two patients is not available.

Model performance for stage 1 – bladder wall, including tumor boundaries when contiguous with the wall segmentation

For stage 1 (bladder segmentation), BLA-T-Seg demonstrated consistently high segmentation accuracy across subsets, with mean DSC scores ≥ 0.85. The best DSC scores were observed in S4 (0.932 ± 0.004), followed by S3 (0.920 ± 0.004), S5 (0.919 ± 0.004), reflecting excellent performance across these subsets. S2 also demonstrated excellent performance (0.905 ± 0.007) with slightly increased variability. While S1 did not achieve excellent performance, the performance was still good (0.848 ± 0.020), with a wider performance range consistent with greater case complexity and inter-patient heterogeneity (Fig. 3A–C).

Fig. 3.

Quantitative assessment of bladder segmentation performance across five subsets: S1, comprising internal pre-treatment T2-weighted images and S2, comprising internal post-treatment T2-weighted images from muscle invasive cancer; S3, comprising internal non-muscle invasive bladder cancer T2-weighted images; S4, comprising external pre-treatment T2-weighted images, and S5, comprising external post-treatment T2-weighted images from muscle invasive bladder cancer. (A) Mean Dice similarity coefficient (DSC) scores with 95% confidence intervals. (B) Boxplots showing the distribution of DSC scores . (C) Violin plots illustrating the density and variability of DSC scores. (D) Radar plot segmentation performance between the internal (United States) and external (Italian) sites. (E) Heatmap summarizing subset-wise mean segmentation metrics.

Cross-site analyses demonstrated minimal performance variation, with institution-level differences in mean DSC scores and complementary metrics (IoU, precision, recall, F1-score, accuracy) typically within 0.01–0.02, underscoring strong generalizability and balanced segmentation behavior (Fig. 3D–E).

Representative examples across training, validation, internal test, and external test sets illustrate accurate contour delineation and strong spatial agreement with expert annotations (Fig. 4, Stage 1 panel), with additional metrics reported in the supplementary material (Figure S1).

Fig. 4.

Representative BLA-T-Seg segmentation results across five subsets: S1-S3, comprising internal study sample of non-muscle invasive bladder cancer and S4-S5 comprising external study sample of muscle invasive bladder cancer. Stage 1 is focused on bladder segmentation, while stage 2 is focused on tumor segmentation. From top to bottom, rows show the original T2-weighted slice, the ground truth bladder mask, the predicted bladder mask after post-processing, and the overlay of the predicted segmentation on the original image.

Model performance for stage 2 – tumor segmentation

For stage 2 (tumor segmentation), BLA-T-Seg performance remained high across subsets, reflecting reliable tumor localization despite the inherent complexity of tumor boundary delineation. The best DSC scores were observed in the S1–S3 combined group, with a mean DSC score of 0.805 ± 0.015, indicating good performance. S4 and S5 demonstrated moderate performance (0.741 ± 0.017 and 0.767 ± 0.018, respectively) with wider interquartile ranges, consistent with greater anatomical and imaging heterogeneity in these subsets that can influence tumor boundary definition (Fig. 5A–C).

Fig. 5.

Quantitative evaluation of tumor segmentation performance across five subsets: S1, comprising internal pre-treatment T2-weighted images and S2, comprising internal post-treatment T2-weighted images from muscle invasive cancer; S3, comprising internal non-muscle invasive bladder cancer T2-weighted images; S4, comprising external pre-treatment T2-weighted images, and S5, comprising external post-treatment T2-weighted images of muscle invasive bladder cancer. Of note, S1-S3 were combined for Stage 2 training because they represent internal cohorts from the same institution. Only testing cases from S1 were included, and S2–S3 were used as withheld test samples in Stage 1 (not used for training), ensuring that all cases were independent of prior training data while providing a sufficiently large cohort for tumor segmentation. (A) Mean Dice similarity coefficient (DSC) score with 95% confidence intervals for tumor segmentation. (B) Boxplots showing the distribution of Dice scores across patients. (C) Violin plots illustrating the density and variability of DSC scores, highlighting subset-wise performance dispersion. (D) Radar plot for segmentation performance between the internal (United States) and external (Italian) sites. (E) Heatmap summarizing subset-wise mean tumor segmentation metrics. **Reported results are within the patient’s threshold tumor volume > 0.5 cm³.

Cross-site analyses demonstrated closely aligned performance between sites, with differences across mean DSC scores and related metrics typically within 0.02–0.04, supporting robust generalizability (Fig. 5D). Across complementary metrics, S1–S3 consistently achieved the strongest performance, followed by S5 and S4; however, accuracy remained uniformly high (> 0.95) across all subsets. No systematic failure patterns were observed, suggesting that performance variations were subset-specific and driven by tumor complexity rather than model limitations (Fig. 5E).

Fig. 4 (Stage 2 panel) shows representative tumor segmentation examples across training, validation, internal test, and external test sets, demonstrating accurate contour delineation and strong spatial agreement with ground truth annotations across diverse clinical settings. Additional quantitative results are provided in the supplementary material (Figure S2).

Model performance benchmarks

Bladder segmentation performance varied markedly across standard architectures (Table 2). DeepLabv3 + showed the weakest performance (DSC = 0.22), whereas the Swin Transformer achieved the highest accuracy (DSC = 0.56), outperforming other CNN-based approaches. In contrast, tumor segmentation performance was uniformly limited across all evaluated standard architectures, with mean DSC values below 0.20. Collectively, these results highlight substantial variability in bladder region segmentation performance across existing models and the consistently challenging nature of tumor-specific segmentation.

Table 2.

Segmentation performance of BLA-T-Seg benchmarked against existing state-of-the-art bladder segmentation methods on study subset S1, as assessed using the Dice similarity coefficient (DSC).

Model Name	DSC (Mean, SD)
Attention U-Net	0.11, 0.08
Deeplabv3Plus	0.26, 0.22
SegformerForSemanticSegmentation	0.10, 0.09
UNet	0.02, 0.03
unet3plus	0.006, 0.0008
UNetASPP	0.16, 0.16
SwinTransformer	0.56, 0.17
UNETR	0.03,0.25
Swin UNETR	0.04,0.15
MRI-CORE	0.12, 0.09
BLA-T-Seg	0.84, 0.14

Discussion

To address unmet clinical and technical challenges in accurate and automated tumor segmentation on bladder MRI, this study developed BLA-T-Seg, based on an adapted SAM. To our knowledge, BLA-T-Seg represents one of the most rigorously validated frameworks for automated bladder and tumor segmentation. A key strength of this study is its rigorous multi-site validation strategy, including independent held-out testing and external evaluation on geographically distinct datasets acquired with heterogeneous scanners and imaging protocols. The anatomically constrained two-stage design is particularly well suited to bladder cancer, where tumors exhibit substantial morphological heterogeneity and complex contours, by restricting tumor prediction to a biologically plausible space and thereby improving structural consistency and reproducibility. BLA-T-Seg demonstrated moderate-to-excellent segmentation accuracy across varied internal and external subsets with mean Dice scores of 0.85–0.93 and 0.74–0.81 for bladder and tumor segmentation, respectively. Compared with 10 SOTA on subset S1, BLA-T-Seg had the highest DSC score of 0.84, over next best-performing model, Swin Transformer (DSC, 0.56). BLA-T-Seg enabled accurate and generalizable bladder wall and tumor segmentation, outperforming SOTA alternatives.

Existing bladder studies using AI have typically focused on a single task such as segmentation often relying on manual or semi-automated tumor delineations that limit scalability and reproducibility [30,31]. Recently, Yu et al. [31]used a deep learning-based model to improve the segmentation accuracy of the bladder walls and tumor on 154 T2-weighted slices achieving DSCs of 98% for the inner wall, 82% for the outer wall, and 87% for tumor in the test set. This study involved a single-center sample without external validation. Moribata et al. [30] proposed a modified U-Net architecture for bladder cancer segmentation using diffusion-weighted (DW) images (b = 0 and 1000 s/mm²), apparent diffusion coefficient (ADC) maps, and multi-sequence MRI as model inputs from a two-center, multi-vendor retrospective sample of 170 patients, the model achieved its best performance, with a mean DSC of 0.81. DW-MRI is used primarily for treatment response assessment and lesion conspicuity, has lower spatial resolution and is more susceptible to artifacts, making it most appropriate as a complementary input rather than a primary segmentation reference. Gumus et al. [32] investigated bladder tumor segmentation using three deep learning architectures, namely U-Net, MAnet, and PSPnet applied to multiparametric MRI, in a small sample of 53 bladder cancer patients. They reported that MAnet with a hybrid cross-entropy and Dice loss function achieved the highest DSC scores across ADC derived from DW-MRI, T2-weighted, and T1-weighted sequences. Trigui et al. [33] segmented the bladder wall in 35 patients using the LevelSet algorithm reaching a DSC of 0.83. These studies remain task-specific and methodologically fragmented, addressing isolated objectives such as tumor segmentation, or radiomics reproducibility, often within relatively small cohorts and without external validation.

Overall, accurate bladder and tumor segmentation may have important clinical implications for treatment planning and therapeutic decision-making in bladder cancer. Automated and reproducible delineation of tumor regions on mpMRI could improve quantitative assessment of tumor burden, local invasion characteristics, and longitudinal treatment response and outcome, particularly in patients undergoing neoadjuvant chemotherapy prior to surgery. Our proposed BLA-T-Seg framework may also support MRI-guided bladder-preservation pathways by enabling more consistent evaluation of residual disease following systemic therapy. Furthermore, the multi-site generalizability demonstrated in this study suggests potential applicability in diverse clinical settings, supporting integration into AI-assisted radiology workflows and decision-support systems.

There are a few limitations of the study. First, although the proposed two-stage BLA-T-Seg framework demonstrates robust and generalizable automated segmentation performance, the generated contours are not intended to replace expert interpretation. Radiologist review and confirmation remain necessary, particularly in real-time clinical settings where complex tumor morphology, post-treatment changes, or atypical anatomy may require manual refinement to ensure diagnostic and therapeutic reliability. In our tumor segmentation failure-case review, underperformance was observed primarily in tumors with volumes < 0.5 cm³, where the model occasionally underestimated or missed subtle lesions because of limited lesion conspicuity, weak tumor–bladder wall boundaries, and partial-volume effects. Representative examples of these underperforming cases are provided in Supplementary (Figure S3). Second, while this study advances beyond prior task-specific and single-institution approaches through rigorous multi-cohort validation, it remains focused on segmentation as a foundational task. Downstream clinical integration, such as real-time interaction with radiologist-in-the-loop, adaptive correction during reporting, or prospective evaluation of workflow efficiency and clinical decision support, was not assessed.

Future studies incorporating radiologist-in-the-loop validation and prospective clinical deployment are essential to establish real-world clinical utility and promote adoption. The framework needs to be validated and include automated volumetric quantification of tumor volumes and supporting standardized longitudinal monitoring as needed. In parallel, inference efficiency optimization and PACS- and cloud-integrated deployment are required.

In conclusion, the BLA-T-Seg framework demonstrated accurate, stable, and generalizable bladder wall and tumor segmentation across internal, held-out, and multi-site external samples, with performance variations primarily driven by anatomical and imaging complexity rather than model limitations. Compared against the 10 SOTA, the BLA-T-Seg substantially outperformed existing methods, underscoring its robustness and suitability for challenging bladder tumor delineation in heterogeneous clinical settings. With this two-stage framework, by explicitly modeling the hierarchical anatomy of the bladder, this structured, context-aware design advances beyond prior single-stage or unconstrained methods, leveraging organ-level information to improve tumor boundary delineation and suppress false-positive predictions outside the region of interest.

Abbreviations:

AI

artificial intelligence

CNN

convolutional neural networks CNN

DSC

Dice similarity coefficient

IoU

intersection over union

ROI

region of interest

SAM

Segment Anything Model

Appendix A. Supporting information

Supplementary data associated with this article can be found in the online version at doi:10.1016/j.ejrai.2026.100107.

Data sharing statement

MRI data generated by the authors or previously analyzed for clinical response assessment are available (USA cohort S1–3 [17] and Italy S4–5 [18]). The auto segmentation code and trained weights will be made available to the research community upon request under a Material Transfer Agreement (MTA).

Data availability

Data will be made available on request.