Incremental Instance Segmentation for the Gleason Tissues Driven Prostate Cancer Prognosis

Abstract

Prostate cancer (PCa) is the second most commonly diagnosed cancer in men and the fifth-highest cause of death globally. Early-stage prostate cancer is frequently asymptomatic and has an indolent course, requiring active observation. Early detection and recognition of Gleason tissue can help handle the PCa spread. Therefore, many deep learning-based systems have been proposed by researchers in order to screen the PCa. Moreover, acquiring such large-scale, well-annotated data can improve the performance of screening and detecting PCa. However, this process is typically challenging and impractical. This paper addresses this issue by proposing a novel knowledge distillation-driven instance segmentation framework. This approach is fused with incremental few-shot training and allows the traditional semantic segmentation models to grade the PCa utilizing instance-aware segmentation, along with the extraction of correlated samples of the Gleason tissue patterns. Furthermore, the proposed approach has been validated on a dataset that contains around 71.7M whole slide image patches. Our approach has outperformed the state-of-the-art models by 2.01% in terms of mean IoU and 9.69% in terms of F1 score for the extraction of Gleason tissue instances and grading PCa, respectively.