The PD-L1 Scoring Problem PD-L1 tumor proportion score (TPS) - the percentage of tumor cells with positive PD-L1 membrane staining - is the primary biomarker determining immunotherapy eligibility in NSCLC. Manual TPS scoring by pathologists is subjective, time-consuming, and shows significant inter-observer variability, especially near the clinically critical 1% and 50% thresholds.
Why Weakly Supervised Learning Most automated TPS systems use fully supervised deep learning requiring extensive cell-level annotations - pathologists must manually label individual tumor cells in thousands of image patches. This annotation burden is expensive and impractical at scale. Weakly supervised learning using only slide-level TPS labels eliminates this bottleneck.
MiLT: Multi-Instance Learning for TPS This study introduces MiLT (Multi-Instance Learning for TPS), an AI tool that predicts TPS from whole slide images (WSIs) using only slide-level labels during training. MiLT divides each WSI into patches, aggregates patch-level features through attention pooling, and predicts a continuous TPS value without requiring cell annotations.
Clinical Context The model was developed and validated on 439 internal samples (Renji Hospital and Shanghai Chest Hospital) and 104 external test samples, all stained with the Dako 22C3 PD-L1 kit - the most widely used companion diagnostic for pembrolizumab in NSCLC.
Tumor Extraction Module Before TPS prediction, a dedicated tumor detection module using MobileNet-V2 (pre-trained on ImageNet) classifies 256x256 pixel patches as tumor vs. non-tumor. This step ensures the downstream TPS model processes only relevant tumor regions, reducing noise from stroma, necrosis, and immune infiltrates that would confound PD-L1 scoring.
MIL Framework Design Each WSI is divided into 256x256 pixel patches at 20x magnification. Extracted tumor patches are randomly organized into bags of 200 patches each (100 bags per WSI). ResNet18 extracts a feature vector from each patch, an attention mechanism assigns importance weights to each patch, and a distribution pooling filter aggregates weighted features for final TPS prediction.
Attention Visualization A key interpretability feature: MiLT generates attention probability maps showing which spatial regions of the WSI were most influential for the TPS prediction. High-attention regions correspond to areas with high PD-L1 positive tumor cell density - providing pathologist-interpretable explanations alongside quantitative scores.
TPS Calculation and Output The model outputs a continuous predicted TPS value on a 0-100% scale, binned into 10% increments matching pathologist reporting conventions (0, 1-9, 10-19, 20-29, etc.). This granular output supports clinical decisions at multiple TPS thresholds rather than just binary positive/negative classification.
Intraclass Correlation Coefficient of 0.960 MiLT achieved an ICC of 0.960 (95% CI: 0.950-0.971) against pathologist TPS scores across the combined internal and external cohorts. This is considered excellent agreement and exceeds the ICC values reported for inter-pathologist agreement in most published studies, suggesting MiLT is more consistent than human scoring.
Robust External Validation The high ICC was maintained in the 104-sample external test cohort from a different hospital, demonstrating that MiLT generalizes across institutions, scanners, and pathology preparation workflows. This cross-institutional robustness is critical for real-world deployment.
Consistent Performance Across TPS Ranges Critically, MiLT maintained high agreement not just at clearly high or low TPS values but also at borderline scores near the clinical thresholds (1% and 50%). Agreement in borderline cases is the most clinically important performance dimension since these are the cases where misclassification directly changes treatment.
Annotation Efficiency Achieved While fully supervised systems require cell-level annotations on millions of cells, MiLT was trained using only slide-level TPS labels provided by pathologists as part of routine diagnostic reporting. This dramatic reduction in annotation cost makes large-scale model training and continuous retraining feasible.
Reducing Inter-Pathologist Variability MiLT's ICC of 0.960 exceeds typical inter-pathologist ICC values, meaning two pathologists using MiLT as an AI second opinion would show better agreement than two pathologists scoring manually. Deploying MiLT as a standardized scoring tool could significantly reduce the current variability that leads to inconsistent treatment decisions.
Scalable PD-L1 Testing Infrastructure As immunotherapy becomes standard of care across multiple cancer types, the volume of PD-L1 tests is growing. MiLT's automated pipeline can process WSIs without pathologist time investment - potentially providing immediate PD-L1 scores upon slide digitization, dramatically reducing turnaround time.
Decision Support at Borderline Cases The most valuable clinical role for MiLT is at borderline TPS values (near 1% or 50%) where pathologist uncertainty is highest. Providing a quantitative AI score with confidence intervals in these cases could flag disagreements for second pathologist review, creating a safety net for high-stakes decisions.
Quality Control and Audit MiLT can serve as a quality control tool in pathology laboratories - if a pathologist's manual score deviates substantially from MiLT's prediction, it triggers review. Over time, analysis of discordance patterns could identify systematic scoring biases or problematic slide preparation factors.
Single PD-L1 Platform Tested MiLT was developed and validated only on the Dako 22C3 PD-L1 assay. Clinical practice involves multiple approved PD-L1 antibody clones (SP142, SP263, 28-8) with different staining characteristics. Expanding MiLT to handle multiple platforms requires separate training or cross-platform adaptation.
Limited Histological Diversity The validation cohort included both lung adenocarcinoma and squamous cell carcinoma but the sample sizes for each subtype were not separately reported. Whether MiLT performance is consistent across histological subtypes with different PD-L1 staining patterns should be assessed.
Combined Positive Score (CPS) Not Addressed MiLT predicts TPS (tumor cells only) but some indications require CPS (tumor + immune cells). A separate module for immune cell detection and CPS calculation would be needed to fully replace manual PD-L1 scoring for all approved drug indications.
Future Directions Multi-platform validation across all PD-L1 antibody clones, extension to CPS scoring, prospective clinical studies demonstrating that MiLT-guided treatment decisions produce equivalent or better outcomes than pathologist-guided decisions, and regulatory submission studies are the key next steps toward clinical adoption.