TransPBMIL: Transformer-Based Weakly Supervised Prognostic Prediction in Ovarian Cancer with Pseudo-Bag Strategy

Ovarian cancer presents a notable health concern characterized by its unfavorable prognosis and elevated mortality rates in the female population. Accurate prognostic assessment is essential for tailoring treatment strategies and improving patient outcomes. Analysis of histopathological whole-slide images is the gold standard for pathological diagnosis, which contains rich phenotypic and molecular information. Multiple instance learning methods have been the dominant approach for processing megapixel whole slide images. However, the methods adopt the one image as a bag strategy, which will contain many noisy tiles leading to model overfitting during training. To mitigate the above situation, we propose a transformer-based multi-instance learning framework with a pseudo-bag strategy (TransPBMIL) for predicting overall survival within 3 years of ovarian cancer patients using pathological images. Extensive studies on multiple cancer prognostic datasets demonstrate the superiority of TransPBMIL.