Automatic Prostate Cancer Detection On Multi-Parametric Mri With Hierarchical Weakly Supervised Learning

Abstract

Multi-parametric MRI (mp-MRI) is one of the most commonly used non-invasive methods for prostate cancer (PCa) diagnosis. In recent years, computer aided diagnosis (CAD) for PCa on mp-MRI based on deep learning techniques has gained much attention and shown promising progress. The key for the success of deep learning based PCa diagnosis is to obtain a large amount of high quality PCa region annotation on mp-MRI such that the network can accurately learn the large variation of PCa lesions. In order to precisely annotate the PCa region on mp-MRI, the pathological whole mount data of the patient is normally required as reference, which is often difficult to obtain in real world clinical situations. Therefore, we are motivated to propose a new deep learning based method to integrate different levels of information available in the PCa screening workflow through a multitask hierarchical weakly supervised framework for PCa detection on mp-MRI. Experimental results show that our method achieves promising PCa detection and segmentation results.