Contrasting Axial T2W MRI For Prostate Cancer Triage: A Self-Supervised Learning Approach

Current diagnostic practices for prostate cancer (PCa) suffer from over-diagnosis of indolent lesions and under-detection of aggressive ones. Deep learning (DL) techniques have shown potential in automatizing tasks and helping clinicians. Nevertheless, their success depends on the availability of large amounts of labeled data, which are rarely available in the medical field. Hence, transfer learning using ImageNet has become the de facto approach but it has been shown to be suboptimal for medical images. Contrastive learning is a form of self-supervised learning (SSL) that leverages unlabeled data to produce pre-trained models and has shown promising results on natural images. However, its application to MRI interpretation has been rather limited. In this work, we propose a contrastive approach (SimCLR) to produce models with better initializations for PCa triage. Our results show that linear and end-to-end fine-tuned models trained on our SSL pre-trained representations outperform ImageNet and random initialization.