Self-Supervised Masked Autoencoders for Prostate Cancer Segmentation via Learned Representations

Akshat Tulsani

Keywords: Self-Supervised Learning, Semantic Segmentation, Medical Imaging, Prostate Cancer, MRI, Vision Transformers, Representation Learning

Abstract: Prostate cancer incidence is rising at an estimated rate of 6.7\% annually. MRI guided biopsy of suspected lesions can improve diagnosis but requires precise delineation of the prostate gland and lesions. Manually delineating lesions is labor intensive and remains challenging due to their morphological similarity to surrounding healthy tissue. In this study, we propose a self-supervised learning framework to learn rich anatomical and pathological representations from bi-parametric MRI (bpMRI) scans. Specifically, we leverage a Masked Autoencoder (MAE) architecture trained on multimodal bpMRI to capture context-aware features of the prostate region through a novel lesion-aware masking strategy. The pretrained encoder is then fine-tuned for lesion segmentation using a lightweight decoder augmented with skip connections from the MAE. Our fine-tuning strategy incorporates a balanced dataset and a hybrid loss function to address class imbalance.The proposed approach outperforms state-of-the-art segmentation methods, achieving over a 10\% improvement in Dice score on the held-out test set.

Submission Number: 8