RIEMAE: Riemannian Masked Autoencoder for Classifying Malignant Prostate Cancer Patterns

Prostate cancer (PCa) is the most frequently diagnosed cancer in men and biparametric magnetic resonance imaging (BP-MRI) has emerged as a non-invasive alternative to characterize clinically significant prostate cancer (csPCa) lesions. However, the assessment of BP-MRI is expert-dependent and challenging due to the remarked variability in malignant and healthy observations. Current computational approaches rely on supervised learning tasks that require large annotated datasets for effective training. This work proposes a novel Riemannian Masked Autoencoder (RieMAE) that captures key relationships in BP-MRI regions of interest, leveraging self-supervised learning on symmetric positive definite (SPD) covariance matrices. The proposed RieMAE considers the Riemannian geometric structure of SPD matrices while reconstructing the original pairwise relationships within radiological findings. The resultant embeddings were evaluated for classifying csPCa lesions, achieving an AUC-ROC of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$94.29\pm 0.88$</tex>, surpassing existing supervised methods.