TEWS: Transformer-empowered weakly supervised prediction of immune score and genetic mutations in liver cancer from whole slide image.

Whole Slide Imaging (WSI) plays a crucial role in predicting immune scores by providing detailed cellular and tissue-level insights, thereby enhancing pathological diagnosis accuracy and biomarker detection. Additionally, WSI contributes essential information for personalized treatment decisions. However, the application of deep learning models to WSI is hindered by two major challenges. First, the substantial data requirements, combined with the high cost and labor-intensive nature of sample annotation, limit the availability of well-labeled datasets. Second, the computational demands of these models pose practical constraints for many healthcare institutions. To address these challenges, we propose a weakly supervised deep learning model based on the transformer architecture. Specifically, we design a novel network incorporating the Swin Transformer, which, unlike traditional convolutional networks, emphasizes global feature extraction. This improves the accuracy of pseudo-label assignment during feature embedding. Additionally, we integrate a gated attention pooling mechanism and employ multi-instance learning (MIL), enabling, for the first time, immune score prediction directly from WSIs. Our model was evaluated through 5-fold cross-validation and achieved an area under the curve (AUC) of 0.88 for immune score prediction in liver cancer. Furthermore, it demonstrated strong predictive performance for genetic mutations, achieving an AUC of 0.99 for CUB and Sushi multiple domains 1 (CSMD1) mutations and 0.86 for Tumor Protein 53 (TP53) mutations in liver cancer. These results highlight the potential of transformer-based weakly supervised learning in computational pathology.