Revolutionizing Prostate Segmentation: A Query Adaptive Pseudo 3D Attention Approach

Motivation: Limited volumetric accuracy and inadequate prediction reliability challenge prostate zonal segmentation (PZS) in MRI, restricting clinical trustworthiness in diagnosis. Goal(s): This study presents a novel model for PZS that incorporates volumetric information and robust relevance estimations, aiming to enhance diagnostic confidence. Approach: We introduce a pseudo-3D model that leverages comprehensive attention mechanisms and a query- context feature to achieve prediction and uncertainity aware PZS evaluated on two large MRI prostate datasets. Results: Our model achieved state-of-the-art performance in PZS with a Dice Score of 0.92 in transitional (TZ) and 0.75 in peripheral (PZ) prostate zones. The query-based relevance improved the interpretability by 12%. Impact: This work supports clinicians in confidently interpreting prostate imaging, potentially reducing unnecessary biopsies and facilitatingtimely intervention. It also encourages advancements in relevance-based AI diagnostics, paving the way forenhanced accuracy and interpretability across medical imaging and broader diagnostic applications.