WeakSTIL: Weak whole-slide image level stromal tumor infiltrating lymphocyte scores are all you need

• DOI: 10.48550/arxiv.2109.05892
• Publication Date: 2021-01-01
• Authors: Yoni Schirris; Mendel Engelaer; Andreas Panteli; Hugo M. Horlings; Efstratios Gavves; Jonas Teuwen
• Keywords: Feature (linguistics); Digital Pathology

We present WeakSTIL, an interpretable two-stage weak label deep learning pipeline for scoring the percentage of stromal tumor infiltrating lymphocytes (sTIL%) in H&E-stained whole-slide images (WSIs) breast cancer tissue. The sTIL% score is a prognostic and predictive biomarker many solid types. However, due to high labeling efforts intra- interobserver variability within between expert annotators, this currently not used routine clinical decision making. WeakSTIL compresses tiles WSI using feature extractor pre-trained with self-supervised on unlabeled histopathology data learns predict precise scores each tile bed by multiple instance regressor that only requires WSI-level label. By requiring label, we overcome large annotation required train existing TIL detection methods. show at least as good other methods when predicting score, reaching coefficient determination $0.45\pm0.15$ compared generated pathologist, AUC $0.89\pm0.05$ treating it clinically interesting sTIL-high vs sTIL-low classification task. Additionally, intermediate tile-level predictions are highly interpretable, which suggests pays attention latent features related number TILs tissue type. In future, may be provide consistent stratify patients into targeted therapy arms.