Automatic framework for PD-L1 expression evaluation in Latino patients with non-small cell lung cancer

Abstract

Programmed death ligand-1 (PD-L1) expression is a key biomarker for identifying non-small cell lung cancer (NSCLC) patients eligible for immunotherapy, but its immunohistochemistry assessment is subject to high interobserver variability. This study aimed to develop and validate an automated system based on artificial intelligence (AI) to detect and classify PD-L1-positive cells in digital lung biopsies from Latin American patients with NSCLC. A total of 141 biopsies were digitized, and deep learning models were trained for tissue segmentation and analysis. YOLOv8 was used for cell detection and ResNet50 for classification into four categories (tumor or non-tumor cells, PD-L1 positive or negative). The tumor proportion score (TPS) was calculated using these classifications. The YOLOv8 model achieved 76% precision and a mAP@0.5 of 63.7% in cell detection, whereas ResNet50 reached 85.2% sensitivity in identifying PD-L1-positive tumor cells. The TPS estimated by the system demonstrated 97% concordance with manual evaluation by an expert pathologist. In addition, spatial PD-L1 maps revealed intratumoral heterogeneity, supporting its relevance for immunotherapy decisions. These findings show that automated PD-L1 quantification using AI is feasible and accurate, improves diagnostic reproducibility, and supports personalized treatment decisions, especially in resource-limited settings.

Introduction

Lung cancer accounts for approximately 12.4% of all cancer cases globally, with an estimated 2,480,675 new cases and 1,817,469 deaths reported in 2022. In Latin America and the Caribbean, it represents 6.8% of all cancer diagnoses, corresponding to 104,952 new cases and 90,550 deaths, with a mortality rate of 12.1%.¹

Immunotherapy has emerged as a promising approach for lung cancer, eliciting robust and durable anti-tumor immune responses that can significantly improve survival by enhancing immune recognition and elimination of cancer cells.2, 3, 4 In non-small cell lung cancer (NSCLC), programmed death ligand-1 (PD-L1) expression is a key immune evasion mechanism mediated through the PD-1/PD-L1 axis. Upon T-cell activation via TCR/MHC peptide recognition, PD-L1 binds to the inhibitory PD-1 receptor, leading to lymphocyte inactivation and immune suppression.5, 6 PD-L1 expression is routinely assessed by immunohistochemistry (IHC) and quantified as the tumor proportion score (TPS), defined as the percentage of viable tumor cells exhibiting membranous PD-L1 staining. TPS is currently the only approved biomarker for identifying NSCLC patients eligible for immune checkpoint inhibitor (ICI) therapy.3, 4

Despite its clinical relevance, TPS evaluation is highly subjective, as it relies on the visual estimation of the pathologist, which introduces significant inter- and intravariability.⁷ Furthermore, the use of various testing systems based on particular PD-L1 antibody clones (e.g., 22C3, SP142, 28-8, SP263), different analytical platforms (e.g., Ventana/Roche, Basel, Switzerland; DAKO/Agilent, Santa Clara, CA, USA), the chosen immunotherapeutic agent (e.g., pembrolizumab, nivolumab, durvalumab, and atezolizumab), and the test's validity based on tumor type further complicates assessment.8, 9 As a result, the reproducibility of PD-L1 scoring across different antibodies and evaluators has been extensively studied, showing variable levels of comparability, interchangeability, and agreement.5, 10

TPS quantification in NSCLC can be conceptualized as a three-step process: (1) identifying individual cells within the image, (2) classifying each as a PD-L1–positive tumor cell, PD-L1–negative tumor cell, or excluded, and (3) calculating the proportion of PD-L1–positive tumor cells relative to the total tumor cell count. These steps make TPS an ideal candidate for automation using deep learning (DL) techniques, which have demonstrated high performance in both cell localization and classification tasks. Automating TPS assessment not only streamlines the workflow but also enables the exploration of novel spatial biomarkers derived from tumor tissue analysis.⁴

DL-based approaches provide a scalable solution by automating the entire TPS evaluation process, from detection and classification to quantification. Whereas promising results have been reported, most studies to date have been conducted in high-income settings, using datasets with limited ethnic and geographic representation.4, 11, 12 To date, no studies have implemented cell-level DL models for automated PD-L1 assessment in Colombian or broader Latin American cohorts. Addressing this gap, an AI-based automated system was developed and validated to assess PD-L1 expression in digitized biopsies from Colombian NSCLC patients. This system is designed to improve diagnostic precision, reduce observer variability, and facilitate the integration of AI into oncological pathology workflows. Beyond TPS calculation, the system enables morphological analysis of tumor cells, supports therapeutic decision-making, and provides a foundation for creating digital databases for longitudinal and multicenter research, advancing personalized cancer diagnosis and care.

Material and methods

This study employed a methodological framework combining advanced DL techniques and artificial intelligence (AI) to automate the identification of the PD-L1 biomarker in histological samples. Digital image management, including storage, processing, and analysis of the images using analytical models, was implemented on the Amazon Web Services (AWS) platform, leveraging resources such as Amazon S3 for storage and Amazon SageMaker for training models on GPU-optimized instances designed for DL tasks.²¹ Model development and evaluation were conducted using Python (version 3.11.9).²²

The components of the proposed methodological framework are described in detail below.

Study cohort

A total of 141 biopsies from Colombian patients with NSCLC were analyzed. The cohort included a spectrum of NSCLC histological subtypes. Among adenocarcinomas, all major invasive patterns classification were identified, including acinar, papillary, micropapillary, and solid subtypes. In addition, cases of squamous cell carcinoma were also included.

PD-L1 immunohistochemical staining was performed on 3 μm-thick sections from formalin-fixed, paraffin-embedded tissue blocks, including both biopsies and surgical specimens. The staining used the Dako 22C3 antibody clone on the automated Dako Link 48 platform. The PD-L1 TPS was expressed as the percentage of tumor cells exhibiting membranous staining, ranging from less than 1% to 100%.

For TPS assessment, an expert pathologist visually divided the tumor area into sections at low magnification. In each section, the percentage of viable tumor cells showing positive membranous staining for PD-L1 was estimated. The overall TPS was determined by calculating the proportion of PD-L1–positive tumor cells relative to the total number of viable tumor cells across all evaluated sections. This study was approved by the CTIC Ethics Committee (file number 050432-4-1911D 2024).

Scoring by pathologists

TPS was initially calculated manually by an expert pulmonary pathologist. Only tumor cells were considered in the evaluation, and those exhibiting partial or complete membranous staining at any intensity level were classified as positive. TPS was calculated using the following formula:

TPS = (Number of PD-L1–positive tumor cells/Total number of viable tumor cells) × 100.

Digitization of biopsies

Lung biopsies, previously prepared and stored on pathology slides, were digitized using a Roche Ventana DP 200 scanner. The resulting images were then stored in a database operating on AWS infrastructure, specifically within an S3 bucket datalake.

At the top of the slide in Fig. 2, a control tissue biopsy, typically obtained from the patient's tonsil, is placed to serve as a reference, ensuring the proper execution of the histopathology procedure. At the bottom, the lung tissue sample, prepared using the same protocol, is located. During image acquisition, the scanner captures multiple focal planes at different depths to internally determine the optimal focus. However, the exported .tif file used for analysis corresponds to a single 2D image (i.e., a single optical plane) at maximum resolution selected by the scanning system. This approach enables precise visualization of tissue-rich areas while ensuring consistency and reproducibility in downstream analysis.(See Fig. 1 .)

Fig. 2.

Pathology slides with lung biopsies and control tissues.

Fig. 1.

Proposed methodology for PD-L1 automation.

In total, 141 lung cancer biopsies were digitized. Each image averaged 800 MB in size, with maximum resolution reaching approximately 71,700 × 43,200 pixels. To expand the dataset and enhance sample diversity, an image partitioning strategy was applied: each biopsy was segmented into multiple subsamples.²⁰ This approach maximized information extraction from each sample, mitigated dataset size limitations, and supported robust training of DL models. Approximately 300 images were generated for cell classification model development. Supplementary Fig. 1 shows slides from different patients.

Identification of control and case tissue

A DL model based on the YOLO (You Only Look Once) architecture was implemented to identify and classify tissue types in the digitized images. YOLO, developed by Redmon et al. (2016),¹⁶ is an advanced object detection model that utilizes convolutional neural networks to locate and classify multiple objects within an image in real time. In this study, the model was configured to identify and classify two specific categories: control tissue and lung tissue, generating bounding box coordinates that encapsulate the areas of interest. The model implementation process included an initial phase of manually selecting and labeling the areas of interest within each sample image. The images were divided into two sets: one for training and another for validation, enabling supervised training and subsequent performance evaluation. The MakesenseAI tool was used for labeling, offering advantages such as being open-source, accessible under the GPLv3 license, and operable without prior installation, as it is accessible directly from a browser. Additionally, this tool supports multiple label types, including rectangles, lines, points, and polygons, and allows export in formats such as YOLO, VOC XML, VGG JSON, and CSV.

The training set was used to adjust and calibrate the model parameters, enabling YOLO to detect and classify the tissues of interest accurately. Meanwhile, the validation set was used to evaluate the model's performance using metrics such as accuracy, sensitivity, and specificity, thereby ensuring reliable results in detecting both control tissue and lung tissue.

Identification of cells within the tissue

Once the lung tissue regions were identified, individual cells within these areas were detected using YOLOv8, an advanced and optimized version of the YOLO architecture known for its high precision and speed in object detection.¹⁷ YOLOv8 was adapted and trained specifically for histological images, using a dataset annotated by an expert pathologist. The annotation process was performed with the MakeSenseAI tool,¹⁸ which allowed precise labeling of cells into four categories: tumor-positive, tumor-negative, non-tumoral-positive, and non-tumoral-negative.

For this stage of the workflow, however, YOLOv8 was configured to detect all cells present, irrespective of their specific category. Due to the high resolution of the digitized images, it was necessary to divide each into smaller fragments (subsamples) with a maximum size of 185 × 185 pixels. This partitioning addressed the computational demands of YOLOv8, improved detection accuracy, and ensured a more systematic analysis. Each original image was systematically fragmented, generating multiple subsamples that were processed by the trained model.

The dataset used to train YOLOv8 was divided into 90% for training and 10% for testing to evaluate model performance. During training, key parameters such as learning rate, batch size, and number of epochs were optimized to achieve a balance between precision and computational efficiency. Additionally, data augmentation techniques, including image rotations, scaling, and adaptive anchor adjustments were applied to improve robustness against variability in tissue appearance.

This strategy enabled thorough and accurate detection of all cells within the lung tissue, creating a foundation for subsequent classification into the four predefined categories. The cell tagging process used for annotation is shown in Supplementary Fig. 2.

Classification of detected cells

The detected cells were classified into four categories using ResNet50, a deep convolutional neural network widely recognized for its effectiveness in image classification tasks.¹⁹ The four predefined categories were: tumor-positive, tumor-negative, non-tumoral-positive, and non-tumoral-negative. For this stage, the cells detected by the YOLOv8 model were cropped from the original subsamples, producing a dataset of 304 high-quality individual cell images.

These cells were extracted from 141 independent biopsies, ensuring representation across multiple patients and tissue regions. Cell selection was guided by expert pathologist annotation, prioritizing morphological diversity and inclusion of all four predefined classes. Although class distribution was not fully balanced, care was taken to avoid over-representation of any single category.

This dataset was divided into a 70% training set and a 30% test set to allow robust evaluation of model performance. During ResNet50 training, key parameters such as the learning rate and number of epochs were tuned for optimal performance. The model's classification accuracy was assessed using a confusion matrix, which provided metrics including accuracy, recall, and precision. These metrics highlighted patterns of misclassification and potential overlap between categories, informing further model refinement. Combining YOLOv8 for detection and ResNet50 for classification enabled both reliable cell identification and accurate categorization, contributing to a detailed analysis of lung tissue and PD-L1 biomarker expression.

Calculation of the tumor proportion score (TPS)

Based on the results of cell detection (identification of cells within the tissue) and classification (Classification of detected cells), the number of PD-L1–positive and PD-L1–negative tumor cells was determined for each subsample. These counts were then aggregated to calculate the TPS for the entire biopsy using the formula:

$${\mathit{TPS}}_{\mathit{Sub}-\mathit{image}}=\frac{\text{Total}\mathrm{PD}-\mathrm{L}1-\text{active cells}}{\text{Total}\mathrm{PD}-\mathrm{L}1-\text{active cells}+\text{Total}\mathrm{PD}-\mathrm{L}1-\text{inactive cells}}.$$

Calculation for the entire sample:

$${\mathit{TPS}}_{\mathit{Sample}}=\frac{{\sum }_{\mathit{img}=0}^{\mathit{img}=n}\text{Total}\mathrm{PD}-\mathrm{L}1-\text{active cells}}{{\sum }_{\mathit{img}=0}^{\mathit{img}=n}\text{Total}\mathrm{PD}-\mathrm{L}1-\text{active cells}+{\sum }_{\mathit{img}=0}^{\mathit{img}=n}\text{Total}\mathrm{PD}-\mathrm{L}1-\text{inactive cells}}.$$

Statistical analysis

Model performance in segmenting, detecting, and classifying PD-L1–expressing tumor cells was evaluated using the following metrics:

Precision: Exclusive to classification models, this metric calculates the proportion of correct predictions over the total positive predictions made by the model.²³ Precision is calculated as the ratio of true positives (TPs) to the sum of TPs and false positives (FPs):

$$\mathit{Precision}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}.$$

Recall (Sensitivity): Also exclusive to classification models, recall indicates the model's ability to correctly identify all relevant instances. It is calculated as the ratio of TPs to the sum of TPs and false negatives (FN):

$$\mathit{Recall}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}.$$

mAP@0.5 (Mean Average Precision at 0.5 IoU): Used specifically in object detection models, this metric measures the average precision (AP) across different confidence thresholds. An Intersection over Union (IoU) value of 0.5 indicates that at least 50% of the predicted bounding box area must overlap with the actual bounding box area to be considered a match.²⁴ A higher mAP@0.5 indicates greater overall precision in detecting and classifying objects within the image.

Results

Segmentation of lungs and control tissue

The performance of the YOLO-based model in identifying and classifying regions of interest as lung tissue or control tissue is summarized in Table 1. Tissue segmentation achieved a mAP at IoU 0.5 (mAP@0.5) of 94.41%, with a recall of 88.57% and a precision of 94.42%.

Table 1.

Key performance metrics for YOLO.

Metric type	Value (%)
Recall	88.57
Precision	94.42
mAP@0.5	94.41

These results indicate that the model reliably identifies the most relevant regions within the images while maintaining a low rate of false detections. The high recall suggests that the majority of true lung and control tissue areas are correctly detected, whereas the high precision reflects that areas predicted as lung or control tissue correspond accurately to the correct categories. Together, these metrics demonstrate the effectiveness of the YOLO model in distinguishing lung tissue from control tissue in histological slides.

Fig. 3 illustrates an example biopsy prediction, where bounding boxes generated by the model accurately highlight lung and control tissue regions, demonstrating correct inference and classification.

Fig. 3.

The model score for labeled sample and control tissue.

Cell detection

Within the segmented lung tissue regions, an average of approximately 200 primary image partitions, each measuring 1280 × 1555 pixels, were generated per sample. These partitions facilitated expert annotation of regions of interest (ROIs) to ensure proper classification and localization of cells within the images. Subsequently, each partition was further divided into smaller 185 × 185-pixel subsamples to optimize YOLOv8 training and detection performance.

This hierarchical partitioning strategy addressed the computational demands of high-resolution histological images, enabling the model to focus on smaller, manageable regions while preserving the overall tissue context. Subsamples systematically covered the entire tissue area, improving detection accuracy and ensuring comprehensive analysis.

The YOLOv8 model demonstrated robust performance in detecting and localizing individual cells within these high-complexity images, achieving a mAP at IoU 0.5 (mAP@0.5) of 63.7% and an accuracy of 76%. These results confirm the model's ability to identify cells reliably, even in regions of high cellular density. Supplementary Fig. 3 illustrates the complete workflow of image partitioning and subsampling, including representative examples of bounding boxes generated by the trained model to highlight detected cells.

Cell classification

We assessed the ResNet50 model's performance in classifying tumor and non-tumor cells. The goal was to calculate the TPS and predict the potential response to PD-L1 inhibitor-based immunotherapies. A confusion matrix, as shown in Table 2, was generated using an expert pathologist's annotations as a reference.

Table 2.

Confusion matrix of the ResNet50 classifier.

Real/Predicted	e = 0	e = 1	e = 2	e = 3
e = 0	13	0	3	0
e = 1	8	46	0	0
e = 2	2	0	24	0
e = 3	1	0	4	1

The results show that the model achieved an overall accuracy of 82.4% in cell classification. For the clinically most relevant categories involved in TPS calculation, the model achieved an accuracy of 54.2% and a recall of 81.3% for negative tumor cells (e = 0). In contrast, for positive tumor cells (e = 1), it achieved a recall of 85.2%. These findings suggest that the ResNet50 model, combined with cell detection using YOLOv8, can be a good tool for evaluating PD-L1 expression and selecting patients for immunotherapeutic treatments.

Calculation of TPS

Application of the proposed methodology to whole lung biopsy samples demonstrated high concordance with manual evaluation. The TPS calculated by the automated system showed 97% agreement with the TPS determined by the expert pathologist, confirming the reliability of the model.

Beyond quantification, the model also generated detailed spatial maps of TPS distribution within the tissue, revealing notable intratumoral heterogeneity in PD-L1 expression.

Fig. 4A illustrates a biopsy sample with a spatial map of PD-L1–positive tumor cell density, highlighting regions of higher expression. In this example, the overall TPS was 82%, underscoring the clinical relevance of identifying spatial patterns of expression.

Fig. 4.

Spatial analysis of tumor proportion score (TPS) in a lung biopsy sample.

(A) Spatial map showing intratumoral heterogeneity of PD-L1–positive tumor cells, with regions of higher density highlighted. (B) Distribution plot of TPS across the tissue, indicating high PD-L1–positive tumor cell density throughout the sample.

Fig. 4B shows the overall TPS distribution across the tissue, indicating a high density of PD-L1–positive tumor cells throughout the sample. Together, these analyses provide a granular view of PD-L1 expression heterogeneity, offering valuable insights for patient stratification and personalized immunotherapy strategies.

Discussion

This study establishes the clinical utility and technical feasibility of automated PD-L1 biomarker evaluation in NSCLC using AI-based algorithms. The developed method, which combines YOLOv8 for cell detection and ResNet50 for morphological classification, demonstrated high concordance (97%) with manual analysis by an expert pathologist, supporting its potential for clinical implementation. A key contribution of this work is the standardization of the TPS, a critical parameter for guiding treatment decisions involving ICIs, such as pembrolizumab and nivolumab.2, 13

Traditional TPS quantification by IHC is prone to interobserver variability due to subjective stain interpretation, differences in pathologist training, and technical variability among antibody clones and analytical platforms.7, 14 The presence of PD-L1–expressing non-tumor cells, such as immune cells, necrotic cells, and macrophages, further complicates interpretation.¹⁵ Visual assessment is also time-consuming and can result in inconsistent estimates, particularly when evaluating large case volumes. Automation addresses these challenges by enhancing reproducibility, standardizing assessment criteria, reducing human error, and promoting equitable access to accurate diagnoses.

Several previous studies have demonstrated that computational models can achieve performance comparable to expert pathologists in detecting and classifying PD-L1–positive cells, even in small or low-quality biopsy specimens. For example, Huang et al. (2022)¹¹ reported a high concordance (r = 0.964, p < 0.001) between an AI system and human pathologists in TPS evaluation for NSCLC, surpassing interobserver agreement among pathologists (r = 0.944, p < 0.001). Similarly, van Eekelen et al. (2024)⁴ validated an AI tool in a hospital setting, reporting high correlation with manual evaluation (correlation coefficients near 0.9), particularly in cases with PD-L1 expression above 50%. That study also highlighted the ability of AI to reduce variability and improve reproducibility which is critical for patient stratification.

More recently, Plass et al. (2025)¹⁵ confirmed strong agreement (correlation coefficients up to 0.94; weighted kappa >0.85) between AI and pathologists across multiple cancer types, including NSCLC, emphasizing AI's advantage in borderline or ambiguous cases. In line with these advances, the present system achieved acceptable performance in cell detection (YOLOv8: precision 76%, mAP@0.5 63.7%) despite the high complexity and heterogeneity of lung tissue. The ResNet50 model achieved strong classification results for PD-L1–positive tumor cells (85.2% recall), which are clinically decisive for ICI eligibility.

Beyond standard quantification, a novel aspect of this study is the automated generation of spatial distribution maps of PD-L1 within tumor tissue. This visualization allows characterization of intratumoral heterogeneity, an emerging factor in predicting immunotherapy response that is often overlooked when relying solely on average TPS values.³ Identifying regions of high PD-L1 density may facilitate more personalized treatment strategies, such as combining immunotherapy with other targeted modalities.

Several limitations of this study should be acknowledged. First, the annotated cell count used for classification (n = 304) was relatively small for training DL models, which may restrict generalizability. Whereas transfer learning and data augmentation were applied to address this limitation, larger and more balanced datasets are necessary to enhance model robustness. Second, additional validation strategies, including k-fold cross-validation and comprehensive hyperparameter optimization, were not conducted due to their computational demands and the proof-of-concept nature of this study. Additionally, dataset splitting was performed at the cell level rather than the patient level, which could introduce data leakage and potentially overestimate model performance.

Additionally, the study utilized data from a single high-complexity institution, which may constrain external validity. Future research will prioritize the development of multicenter validation frameworks to evaluate model performance across varied clinical environments. Incorporating data from multiple institutions will introduce variability in tissue processing, staining protocols, scanner technologies, and patient demographics, thereby enabling a more comprehensive assessment of model generalizability and robustness. Expanding the dataset in this manner will also support the adoption of patient-level validation strategies, advanced cross-validation methods, and more sophisticated optimization techniques. Moreover, multicenter collaboration will facilitate the use of domain adaptation methods to address interinstitutional variability and enhance model transferability. Concurrently, future studies will assess model robustness across different PD-L1 antibody clones and imaging platforms, and investigate the integration of spatial biomarkers and explainability methods to improve clinical interpretability and predictive accuracy.

In conclusion, this study demonstrates the feasibility of developing an automated, accurate system for PD-L1 evaluation in lung biopsy samples using AI. Integrating this technology into clinical workflows could improve diagnostic quality, reduce interobserver variability, and increase efficiency, particularly in resource-constrained settings. Moreover, it provides a promising foundation for developing novel spatial biomarkers and advancing personalized immunotherapy approaches in lung cancer.

Ethical statements

This study was reviewed and approved by the Kayre Ethics Committee, approval No. 164 Bogotá, Colombia, and was conducted in accordance with the Declaration of Helsinki (2013 revision). Written informed consent was obtained from all patients for participation and for publication of anonymized data and images. Patient confidentiality and anonymity were strictly protected throughout the study.

Funding

This work was supported under an unrestricted grant in a collaboration agreement between Bristol Myers Squibb of Colombia S.A. and CTIC Foundation Cancer Treatment and Research Center Luis Carlos Sarmiento Angulo.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

Dr. Andrés F. Cardona reports grants from Merck Sharp & Dohme, Boehringer Ingelheim, Roche, Bristol-Myers Squibb, and The Foundation for Clinical and Applied Cancer Research – FICMAC., other from Pfizer, Boehringer Ingelheim, Astra Zeneca, MSD, BMS, Celldex, Roche, personal fees from Merck Sharp & Dohme, Boehringer Ingelheim, Roche, Bristol-Myers Squibb, Pfizer, Novartis, Celldex Therapeutics, Foundation Medicine, Eli Lilly and Foundation for Clinical and Applied Cancer Research – FICMAC, outside the submitted work.

Jairo Zuluaga received honoraria as an advisor and participated in the speakers' bureau of Astra Zeneca, Bristol Myers Squibb, and Pfizer. Leonardo Rojas received honoraria as an advisor and participated in the speakers' bureau from Merck Sharp & Dohme, Boehringer Ingelheim, Roche, Bristol-Myers Squibb, Astra Zeneca, and Eli Lilly. Additionally, he was linked and received honoraria as a researcher. Oscar Arrieta reports personal fees from Pfizer, grants and individual fees from Astra Zeneca, grants and individual fees from Boehringer-Ingelheim, personal fees from Lilly, individual fees from Merck, personal fees from Bristol Myers Squibb, grants and personal fees from Roche, outside the submitted work. The other authors declare no conflict of interest.