A Multimodal Geometric Deep Representation to Support Bi-Parametric Prostate Cancer Lesion Classification

Bi-parametric MRI (BP-MRI) sequences are today a PI-RADS standard to characterize clinically significant regions (CSR) without additional contrast parameters. These CSR are characterized according to morphology, microcirculation, and cellular density, but their analysis remains expert-dependent, reporting a significative low specificity. Despite current advances in deep strategies for CSR classification, modeling high textural variability depends on a significant amount of stratified trained data to deal with textural variability. This work introduces a multimodal deep geometrical network that integrates T2WI and (ADC, B-VAL) convolutional branches into a symmetric positive matrix (SPD) embedding. This SPD embedding preserves second-order patterns, which are then propagated through proper geometrical layers to exploit manifold patterns and achieve effective discrimination of CSR patches. The proposed approach outperformed the baseline convolutional architecture, achieving a AUC-RC of 0.86. Interestingly, with only 20% of the labeled lesions (~ 64 labeled samples), the proposed approach achieved a AUC-ROC score of 0.84.