Abstract
Prostate cancer is the second most commonly diagnosed malignancy in men, yet interpretation of multiparametric MRI (mpMRI) suffers from substantial inter-observer variability and depends on scarce expert radiologists. Expert readers follow a hierarchical, coarse-to-fine visual strategy: they first localize the prostate gland and then direct focused attention to tissue within it. Inspired by this biological selective-attention mechanism, we propose a segmentation-guided framework in which a modality-aware 3D Vision Transformer (ViT) first delineates the prostate gland from T2-weighted, ADC, and DWI volumes, and the resulting anatomical mask then constrains downstream cancer detection. Global self-attention models long-range anatomical context, analogous to the integration of peripheral and foveal information in human vision. Detection was evaluated through two pathways: gray-level co-occurrence matrix (GLCM) texture features with classical classifiers, and a ViT classifier. On 200 patients from the public PROSTATEx cohort, the proposed segmentation model achieved a Dice similarity coefficient of [0.905] versus [0.831] for an optimized U-Net. The segmentation-guided ViT achieved an AUC of [0.987] and an accuracy of [0.998] for clinically significant cancer detection. Ablation experiments confirmed that anatomical guidance improved both pathways. These results show that emulating expert visual search strategies yields accurate and interpretable computer-aided prostate cancer diagnosis.