A weakly supervised deep learning framework for automated PD-L1 expression analysis in lung cancer

Abstract

The growing application of immune checkpoint inhibitors (ICIs) in cancer immunotherapy has underscored the critical need for reliable methods to identify patient populations likely to respond to ICI treatments, particularly in lung cancer treatment. Currently, the tumor proportion score (TPS), a crucial biomarker for patient selection, relies on manual interpretation by pathologists, which often shows substantial variability and inconsistency. To address these challenges, we innovatively developed multi-instance learning for TPS (MiLT), an innovative artificial intelligence (AI)-powered tool that predicts TPS from whole slide images. Our approach leverages multiple instance learning (MIL), which significantly reduces the need for labor-intensive cell-level annotations while maintaining high accuracy. In comprehensive validation studies, MiLT demonstrated remarkable consistency with pathologist assessments (intraclass correlation coefficient = 0.960, 95% confidence interval = 0.950-0.971) and robust performance across both internal and external cohorts. This tool not only standardizes TPS evaluation but also adapts to various clinical standards and provides time-efficient predictions, potentially transforming routine pathological practice. By offering a reliable, AI-assisted solution, MiLT could significantly improve patient selection for immunotherapy and reduce inter-observer variability among pathologists. These promising results warrant further exploration in prospective clinical trials and suggest new possibilities for integrating advanced AI in pathological diagnostics. MiLT represents a significant step toward more precise and efficient cancer immunotherapy decision-making.

Keywords: MiLT; PD-L1; TPS; automated scoring; lung cancer.

Figure 1

Consistency of the pathologists and MiLT in the internal and external test cohorts. Scatter plots of TPS-AI vs. TPS-Truth with intraclass correlation coefficient (ICC) in internal test cohort (A) and external test cohort (B). Comparison of Cohen’s kappa values between AI and manual identification based on different cut-off values (1% and 50%) in internal test cohort (C) and external test cohort (D).

Figure 2

Examples of patches and performance of the classification model. (A) Typical patches of tumor areas and other regions. Scale bar: 50 μm. All patches are of the same size. (B) Evaluation metrics of the model’s performance on tumor patch classification, including Accuracy, Precision, Recall, and F1 score. The y-axis represents the metric values ranging from 0.00 to 1.00. (C) Pattern diagram of whole slide images (WSIs) divided into smaller patches of 256 x 256 pixels. Typical examples of tumor patches are magnified for better visualization. In the WSI, tumor patches are displayed, with the tumor regions marked in blue among all tumor patches.

Figure 3

PD-L1 TPS assessments of MiLT and pathologists on different PD-L1 expression levels.The accuracy of TPS scores of AI based on confusion matrix analysis in internal test cohort (A) and external test cohort (B). Comparison of histograms of DL model performance based on PD-L1 expression at different cut-off values (0% - 1% vs. 1% - 49% vs. 50% - 100%) in internal test cohort (C) and external cohort (D).

Figure 4

Original images and the corresponding heatmaps of model visualization. (A, C) The original images stained for PD-L1, with the brown-stained areas representing PD-L1 positive tissues. (B, D) Distribution heatmap for the entire image, set with different TPS thresholds.

Figure 5

The entire workflow of the proposed deep learning framework. The entire workflow consisted of three parts, beginning by cropping the input WSI into patches and extracting tumor patches through the classification module, with patches randomly placed into bags. The core part was the MIL module, where the model took a bag of patches as input and predicted the sample’s TPS in its output. The feature extractor module extracted a feature vector for each patch within the bag. The attention module calculated attention scores based on the feature vectors and assigned weights to the patches. The MIL pooling filter summarized the extracted features into a bag-level representation by estimating the marginal feature distribution. Finally, the bag-level representation transformation module predicted the sample-level TPS. The TPS values inferred by multiple experienced pathologists were used as labels during training.