Abstract Prostate cancer is the second most common cancer in men worldwide. Early detection is critical for reducing deaths and improving patients’ quality of life. Bi-parametric magnetic resonance imaging (bpMRI) has emerged as a potential alternative for supporting the early detection, diagnosis, and screening of prostate cancer. The inherent ambiguity and shape variability of prostate cancer lesions in bpMRI present a major challenge for automated detection systems. To address this, we propose a multitask deep representation that explicitly leverages lesion boundary information to improve localization. Unlike traditional methods that treat localization or segmentation as independent tasks, our approach integrates a segmentation branch that forces the network to learn fine-grained contour details. This synergy results in a more robust feature representation, improving the model’s ability to distinguish clinically significant lesions from surrounding tissue. Our approach was validated on a public dataset of 3343 slices from 1295 bpMRI studies. The proposed model achieved an Average Precision (AP@0.5) of 0.59, a precision of 0.66, and a recall of 0.56. This represents a 22.9% improvement in AP@0.5 compared to a standard localization-only network. Furthermore, compared with PI-CAI reports, our model achieved an AP@0.1 of 0.72 under the commonly used criterion, while the reported AP@0.5 serves as a complementary stricter evaluation within our protocol. In an additional validation using annotations from an expert radiologist as ground truth, the model achieved a precision of 0.79. Our findings indicate that forcing a network to be boundary-aware through segmentation is an effective strategy for improving lesion localization.
González et al. (Mon,) studied this question.