A total of 210 formalin-fixed, paraffin-embedded, anonymized samples from patients diagnosed with gastric cancer were collected from 3DMed Clinical Laboratory (accredited by CAP and CLIA) as model development and internal test cohort in this study. Among these, 100 samples were used to develop the deep learning (DL) models, while the remaining 110 samples constituted a held-out internal test set to evaluate the AI-based CPS prediction pipeline performance. Besides, 98 external samples were obtained from Shanghai Renji Hospital and used as external cohort to test the generalization ability of the AI-based pipeline. All the samples were prepared and stained using the PD-L1 IHC 22C3 pharmDx assay (Dako, Carpenteria, CA, USA) on the Dako Autostainer Link 48 platform, according to the manufacturer’s protocol. After the completion of section staining, all the tissues on the stained sections were scanned and digitized at 20× magnification (0.475 μm/pixel) as WSIs using a KFBIO FK-Pro-120 slide scanner. The exclusion criteria for samples include severe tissue folds/tears, strong nonspecific staining, the presence of large bubble issues, among others. The interpretations of CPS values were performed by two trained pathologists (PD-L1 22C3 assay certified) under double-blinded conditions. To ensure the precision and reliability of the DL model, only WSI samples with concordant diagnoses from two pathologists were retained as ground truth for subsequent comparison with DL model outputs.

To achieve automated prediction of PD-L1 expression in gastric cancer, we constructed an AI algorithm-based prediction pipeline. This pipeline integrates sequential deep learning models operating without manual intervention during testing. Each model was individually trained and validated with corresponding annotated data. All data annotations were performed by pathologists using an in-house developed software (APTime, developed by 3D Medicines Inc.). The fully automated pipeline for CPS prediction, as illustrated in Figure 1C , initiates with tissue localization. It was performed on WSIs using Otsu thresholding on grayscale-converted slides at 0.625 × magnification ( Figure 1A ). Otsu preprocessing significantly enhanced computational efficiency by eliminating redundant patch classification across non-informative background regions. During the subsequent model prediction phase, all processing was conducted at a magnification of 20 × (0.475 μm/pixel). Identified tissue regions were then partitioned into non-overlapping patches with 256 × 256 pixel size. A trained MobileNet-v2 patch classifier then categorized these patches as either tumor-containing or non-tumor-containing. To refine tumor regions identification, patches classified as tumor by the MobileNet-v2 were then processed by a trained U-Net model for pixel-level segmentation of tumor versus non-tumor regions. Only patches exhibiting consensus tumor regions (those classified as tumor by MobileNet-v2 and simultaneously segmented as tumor by U-Net) were retained for subsequent tumor cell analysis. At last, the trained YOLO-based detector performed triple-task recognition: detection of (1) PD-L1⁺ tumor cells, (2) PD-L1⁻ tumor cells in the tumor regions, and detection of (3) PD-L1⁺ immune cells in tumor-containing patches associated non-tumor regions. The final CPS was calculated based on the cellular counts derived from YOLO detection outputs.

To prepare the training dataset of the patch classification model, the 100 WSIs used for model development cohort were divided into 256 × 256 pixel patches. For samples with extensive tumor regions, pathologists selected those representative tumor patches. For samples containing limited tumor regions, we retained as many tumor-containing patches as possible to incorporate into the training data set. These patches were labeled as either tumor-containing patches or non-tumor patches based on the presence or absence of tumor cells. To maintain class balance and enhance model performance during training, we selected a comparable number of non-tumor patches to tumor-containing patches from each sample, resulting in a final dataset comprising 118,715 tumor-containing patches and 119,476 non-tumor patches (including necrotic areas, normal epithelial regions, stromal regions, etc.).

The dataset was partitioned into training, validation, and test subsets at a 6:2:2 ratio, and to enhance model robustness, we use random flip, rotation, and blur to augment the data during training. We employed the MobileNet-v2 architecture, a lightweight convolutional neural network (CNN) optimized for computational efficiency, as the backbone for patch classification (19). The model leveraged transfer learning through ImageNet pre-trained weights, with strategic fine-tuning: only the final seven layers were unfrozen to adapt domain-specific histopathological features while preserving generic pattern recognition capabilities from pre-training. Following the convolutional layers, the architecture includes a flatten layer and a dense layer ( Figure 1B ).

To prepare the training dataset of the tumor segmentation model, representative patches with 1024 × 1024 pixels size from the 100 WSIs in the model development dataset were selected for pixel-level annotation. We constructed a dataset comprising 3,923 image patches. Among these, 1,929 tumor-containing patches were annotated with pixel-level tumor region labels by pathologists, while 1,994 additional patches containing normal tissues or adjacent non-tumorous tissues were incorporated into the training set, serving as background to enhance the model’s ability to distinguish tumor boundaries.

The datasets were randomly split into training, validation and test set in a ratio of 7:2:1. Data augmentation including random flipping and rotation, and hue, saturation and value change were used during training to avoid overfitting and improve accuracy and generalization ability of the model. Tumor segmentation model was built based on a U-Net structure, with Xception-style block, which consists of separable convolution layer to reduce the number of parameters and accelerate inference (20). The model is a symmetric encoder-decoder architecture with skip connection. In the training procedure, labeled patches were further cropped to 512 × 512 size during training. The DL model was trained to simultaneously segment the tumor area and classify the input region as auxiliary loss. Only the output of the tumor area segmentation task was used to predict the tumor region ( Figure 1B ).

To prepare the training dataset of the cell detection model, representative patches (256 × 256 pixel size) containing PD-L1⁺ and PD-L1⁻ tumor cells, or PD-L1⁺ immune cells, were selected from the 100 cropped WSIs used for model development. We finally constructed a dataset comprising 4604 image patches. On these patches, cells were annotated by experienced pathologists using spots with cell tags and were grouped into PD-L1⁺ tumor cells (85,659), PD-L1⁺ immune cells (19,434), and PD-L1⁻ tumors cells (130,512). When annotating PD-L1-positive cells, we labeled cells exhibiting diverse expression intensity levels (including strong, moderate, and weak). For immune cells, we also labeled various morphological forms of both lymphocytes and macrophages.

We built the cell detection model based on the YOLOX (21), which can directly classify, locate, and count the objects on the input patches ( Figure 1B ). In the data augmentation step, the same strategies as U-Net and mosaic and mixup were applied.

CPS is generally calculated by dividing the number of PD-L1 stained cells (including tumor cells, lymphocytes and macrophages) by the total number of viable tumor cells, multiplied by 100. The total formula is shown below:

To ensure precise calculation of the CPS, the following criteria must be rigorously applied: 1) PD-L1-positive tumor cells are defined as tumor cells exhibiting partial or complete linear membrane staining within tumor nests, excluding cells in necrotic areas. 2) PD-L1-positive immune cells should be quantified only if they are located within tumor nests or adjacent supporting stroma and maintain direct spatial proximity to tumor cells (within a 0.5 mm radius).

As cells were grouped into PD-L1⁺ tumor cells, PD-L1⁺ immune cells, and PD-L1⁻ tumors cells by cell detection model, the final formula is shown below:

In clinical diagnosis, pathologists approximately distinguish the tumor cell region from other regions firstly at the lower magnification scale and then zoom into the higher magnification for accurate cell counting, and ultimately render a definitive CPS positive/negative assessment. In this study, the final CPS was calculated using counts of PD-L1⁺ tumor cells, PD-L1⁺ immune cells, and PD-L1⁻ tumor cells detected by YOLOX.

The CPS threshold for PD-L1 positivity is defined as ≥ 1. Based on this cutoff, samples were stratified into two distinct subgroups: PD-L1⁻ samples: CPS < 1; PD-L1⁺ samples: CPS ≥ 1.

This study employed a comprehensive set of evaluation metrics to assess AI model performance, including: Classification metrics (accuracy, precision, recall, specificity, and F1 score); Segmentation metrics (dice coefficient and pixel accuracy); Target detection metrics (Intersection over Union (IoU), Average Precision (AP)). Consistency between AI-calculated CPS values and pathologist assessments was analyzed using confusion matrices and Cohen’s kappa coefficient. The kappa statistic (range: 0 - 1) was interpreted using established clinical benchmarks: slight agreement (0 - 0.2); fair agreement (0.2 - 0.4); moderate agreement (0.4 - 0.6); substantial agreement (0.6 - 0.8); near-perfect agreement (0.8 - 1.0).

All statistical analyses and graphical visualizations were conducted using Microsoft Excel and Python (version 3.9.12), implemented through the PyCharm 2021.3.3 integrated development environment.

For the 210 specimens utilized in model development and internal validation. As detailed in Table 1 , both cohorts demonstrated comparable clinicopathological characteristics. The majority of the samples were surgical resection 79% (166/210), and a minority were needle biopsy and others 21% (44/210). All tumor samples were exclusively collected from the stomach. No statistically significant differences were observed between the two groups (all p-values > 0.05).

To evaluate the PD-L1 expression in the tumor region of a sample, it is essential to accurately localize the tumor area. Considering that patch-level classification models are more efficient in analysis compared to pixel-level segmentation models, and the annotations required for training patch classification models are relatively easier to obtain, we first trained a patch classification model to localize the tumor region. We divided the annotated patches into a training set, a validation set, and an independent test set. The trained model demonstrated high performance on both the validation ( Figure 2A ) and test sets ( Figure 2B ), with accuracy, specificity, and sensitivity all exceeding 97% ( Table 2 ). The trained classification model also effectively distinguished stained necrotic regions from tumor regions ( Figure 2C ), thereby eliminating the impact of necrotic regions on PD-L1 evaluation.

The tumor patches identified by the patch classification model not only encompass tumor regions but also contain partial stroma areas. Therefore, a segmentation model is further required to distinguish between tumor and stroma.

In this study, we trained a segmentation model that can distinguish tumor regions from non-tumor regions, which include necrosis and normal epithelium. The model’s performance was evaluated using dice coefficient and pixel accuracy metrics. On both the validation and test sets, the model demonstrated high segmentation performance. Notably, it effectively identified non-tumor regions, with all metrics exceeding 97% ( Figures 3A, B ). Additionally, the segmentation model can accurately distinguished tumor regions from necrotic areas and normal glandular structures ( Figure 3C ).

Within our PD-L1 expression evaluation pipeline, precise tumor region localization is followed by quantification of PD-L1⁺ tumor cells, PD-L1⁻ tumor cells, and PD-L1⁺ immune cells within tumor-associated regions. To accomplish this, we developed a deep learning-based cell detection model utilizing the YOLO framework for identification of these three cellular phenotypes. Since immune cells can infiltrate into the tumor region, the calculation of the CPS requires integration with the output of tumor segmentation model. This allows us to distinguish PD-L1⁺ tumor cells, PD-L1⁻ tumor cells, PD-L1⁺ immune cells within the tumor regions, and PD-L1⁺ immune cells within the non-tumor regions for final CPS calculation ( Supplementary Figure S1 ).

We evaluated the model’s performance using IoU and AP, with an IoU threshold set at 0.5. The trained cell detection model demonstrated strong performance on both the validation and test sets, achieving AP scores close to 0.900 for true positives (0.889 on the validation set and 0.888 on the test set) ( Table 3 ).

To validate the accuracy of our AI-based pipeline, we assessed the agreement between CPS-AI (AI-derived CPS) and CPS-Doc (pathologist-evaluated CPS) using Cohen’s kappa coefficient in internal and external cohorts. First, we evaluated our pipeline on a held-out internal test cohort (n = 110), which was excluded from model training. To further test generalizability, an independent external cohort (n = 98) was introduced. As shown in Figure 4 , the internal dataset achieved a kappa value of 0.782, demonstrating substantial agreement between model predictions and pathologist-evaluated scores. While the external dataset exhibited a slightly lower value of 0.737, the kappa value remained clinically meaningful, confirming the model’s robustness across diverse datasets.

To further quantify performance, confusion matrices were generated to compare CPS-AI against CPS-Doc. Using CPS-Doc as the reference standard, we evaluated the CPS prediction accuracy of the AI-based pipeline across multiple metrics. In the internal test cohort, AI-based pipeline achieved an accuracy of 0.882 [95% CI = 0.822 - 0.942], sensitivity of 0.964 [95% CI = 0.929 - 0.999], and specificity of 0.800 [95% CI = 0.725 - 0.875], highlighting its strong discriminative capability ( Figures 5A, C ). Performance remained robust in the external cohort, with retained accuracy and high sensitivity ( Figures 5B, C ). These results collectively underscore the reliability and generalizability of the AI pipeline.

In this study, we combined a patch classification model with a region segmentation model to localize tumor areas, where only tumor regions simultaneously identified by both models would proceed to subsequent cell detection. This design offers dual advantages: On one hand, integrating results from both models may enhance the accuracy of tumor region identification. On the other hand, leveraging the higher efficiency of the patch classification model to first roughly localize tumor regions, followed by region segmentation on these pre-selected patches, significantly improves the overall analysis efficiency. Our findings demonstrate that the integrated pipeline incorporating the patch classification model achieved slightly improved consistency with the pathologists compared to the pipeline using only region segmentation with cell detection ( Figure 4 , Supplementary Figure S2 ). Regarding efficiency, the integrated workflow showed 25 - 30% improvement in average processing time per sample ( Table 4 ). Particularly for samples with small tumor areas relative to the whole tissue section, a more than twofold enhancement in processing efficiency was achieved ( Supplementary Figure S3 ).