This paper presents a three-stage deep learning framework for automated PD-L1 Tumor Proportion Scoring (TPS) in non-small cell lung cancer (NSCLC). The core innovation is TPS-Guard, an uncertainty-guided triage system that identifies difficult cases for pathologist review, particularly those near critical clinical thresholds (1% and 50%), where scoring errors can significantly impact treatment eligibility.
TL;DR
The authors developed a three-stage AI framework that automates PD-L1 scoring and uses model ensemble disagreement to triage the most "uncertain" cases to pathologists. The system achieves human-level accuracy and demonstrates that model uncertainty is a strong predictor of both automated error and inter-pathologist disagreement.
What Problem Does This Paper Address?
Immunotherapy selection for lung cancer patients relies heavily on the PD-L1 Tumor Proportion Score (TPS)—the percentage of viable tumor cells showing membranous staining. However, manual scoring is notoriously difficult. Pathologists must distinguish stained tumor cells from "mimics" like macrophages and immune cells, often in heterogeneous tissue samples.
Small errors are especially consequential near the 1% and 50% cutoffs, which determine whether a patient receives specific therapies. While previous AI tools have attempted to automate this, they often lack a mechanism to signal when they are likely to be wrong. This paper addresses the need for a "safety net" that integrates AI efficiency with human expertise.
Methods
The framework follows a logical three-stage sequence:
- Stage 1: Tumor Segmentation: Uses a linear classifier on features from Virchow2 (a vision transformer foundation model) to identify tumor regions and, crucially, exclude macrophage-rich patches that often lead to false positives.
- Stage 2: D2Count-Net: A dual-branch cell counting network. One branch performs point detection (ideal for sparse cells), while the other performs density estimation (better for crowded regions). A "scene classifier" dynamically weights these two outputs based on local cellularity.
- Stage 3: TPS-Guard: This is the triage layer. The authors trained an ensemble of 15 models. Uncertainty is defined as the standard deviation of the TPS estimates across these models. If the models disagree significantly, or if the score sits precariously near a clinical cutoff, the case is flagged for manual review.
Datasets
- Development Cohort (UHCW): 154 whole-slide images (WSIs) from University Hospitals Coventry and Warwickshire, stained with the SP263 assay. Three pathologists provided independent scores to establish a "three-reader mean" reference.
- External Validation Cohort (IGNITE): 98 regions of interest (ROIs) from two European centers, featuring different antibody clones (SP263 and E1L3N) and scanners to test generalizability.
Key Insights and Ideas
- Uncertainty as a Proxy for Difficulty: A major "aha" moment was finding that model uncertainty correlates strongly with inter-pathologist disagreement ($\rho = 0.71$). This suggests that the cases the AI finds "hard" are the same ones humans struggle to agree on.
- Hybrid Counting is Essential: By combining detection and density branches, the model handles the full spectrum of tissue morphology—from scattered tumor cells to dense, solid nests—more robustly than either method alone.
- Triage Improves the Whole System: Referring just the 15% most uncertain slides significantly reduced the Mean Absolute Error (MAE) for the remaining "auto-reported" cases, bringing the AI's performance in line with the most accurate pathologists.
Key Results
- Human-Level Accuracy: On unseen test slides, the model's MAE was 6.96 percentage points, falling squarely within the range of the three expert pathologists (4.24 to 8.29).
- Effective Triage: After referring the 15% most uncertain slides, the MAE on the remaining auto-reported slides dropped to 4.96.
- Zero-Shot Generalization: When applied to the external IGNITE dataset without retraining, the triage system still successfully directed review to higher-error cases, although absolute error was higher on the E1L3N clone (which the model hadn't seen during training).
- Sensitivity at Cutoffs: Referring more uncertain cases raised the sensitivity at the critical 50% cutoff, ensuring fewer high-PD-L1 patients were missed.
Limitations and Future Work
- Clone Sensitivity: The model performed significantly better on the SP263 clone (used in training) than the E1L3N clone. Future work should focus on more diverse training sets to improve "out-of-the-box" performance across different assays.
- Class Imbalance: The training data was imbalanced, with only 6% of annotated nuclei being PD-L1 positive. The authors noted that more positive cell annotations are needed to further refine the counting branches.
- Prospective Validation: This was a retrospective study. The next step is a prospective "reader study" to see how TPS-Guard actually affects pathologist workflow, confidence, and reporting time in a live clinical environment.
Summary updated