Abstract
Soft tissue tumor is a general term for tumors that arise in soft tissues such as fat, muscle, and nerve, and its malignant variants are classified as rare cancers. Rare cancers are more difficult to diagnose pathologically, and in addition, there is a chronic shortage of pathologists in Japan. While core needle biopsy (CNB) is considered technically more difficult than incisional biopsy (IB), its lower invasiveness makes CNB the preferred method for diagnostic confirmation when feasible. In this study, we propose a method for predicting patient survival time using weakly supervised learning on pathology images obtained from CNB. In the proposed method, the pathology images collected from each patient are divided into small regions, and each small region image is labeled with a survival time value corresponding to the original image. Pathological features are extracted from each small region using a pre-trained model (UNI) based on a large-scale pathological image dataset. IB pathology images, which are considered to have relatively high diagnostic reliability, are used as the teacher data, and regression of survival time values is performed on CNB pathological images. From the distribution of predicted values across all small regions, several statistical methods are used to derive the patient-level survival time prediction. As a result, the best mean absolute error (MAE) achieved was 9.48 months.
