Zonal-Aware Prostate Cancer Detection and Classification via Hadamard-Fused Mamba-UNet

Abstract

Accurate clinically significant prostate cancer (csPCa) detection from prostate MRI requires a detailed understanding of the prostate anatomical context. Deep learning approaches have achieved promising results in zonal and lesion segmentation; however, the joint analysis of prostate zonal anatomy and lesion morphology in a unified framework remains relatively underexplored. In this work, we propose ZH-Mamba, a multi-stage zonal-aware csPCa detection framework consisting of two H-Mamba UNets models for prostate zonal segmentation and lesion localization, followed by an anatomy-aware Convolutional Neural Network classifier for csPCa detection. The proposed model leverages a learnable Hadamard Transform-based fusion mechanism to enhance zonal and lesion feature representations without increasing computational complexity. The learned fusion patterns consistently emphasize prostate regions, suggesting that the proposed fusion captures anatomically relevant feature representations. We demonstrate a strong zonal segmentation performance on the PI-CAI dataset, reaching Dice scores of $\textbf{0.95}$, $\textbf{0.85}$, and $\textbf{0.95}$ for whole gland (WG), peripheral zone (PZ), and transition zone (TZ), respectively, providing robust anatomical priors for downstream csPCa detection. Experimental results on an independent internal dataset confirm the generalization capability of the proposed zonal segmentation model. For lesion localization, the H-Mamba UNet achieves a competitive Dice score of $\textbf{0.70}$. Integrating the zonal masks and lesion localization, the proposed ZH-Mamba framework achieves an AUC of $\textbf{0.86}$, Average Precision ($\mathrm{AP}_{\mathrm{Patient}}$) of $\textbf{0.72}$, and F1 score of $\textbf{0.76}$, demonstrating the effectiveness of zonal-aware csPCa detection.