Enhanced prostate zone segmentation with multi-timepoint MRI: a cross-attention augmented 3D U-Net approach

Accurate prostate zone segmentation is an important preliminary step for the diagnosis and treatment of prostate cancer (PCa). While the clinical focus in active surveillance (AS) patients with low-risk or favourable intermediate-risk PCa ultimately lies in the precise segmentation and characterisation of target lesions, prostate zone segmentation remains an important foundational step for accurate diagnosis of PCa. Regular non-invasive monitoring using MRI plays a critical role in AS, where multiple scans are acquired at different time points to track disease progression. However, interpreting subtle PCa cases in AS using MRI remains challenging due to the difficulty in identifying small lesions within the prostate. To address these challenges, more robust deep-learning models for automated prostate zone segmentation have been developed, which helps improve lesion detection for accurate disease monitoring by detecting low-risk cancers progressing to high-risk prompting active treatment. State-of-the-art prostate zone segmentation models are typically trained on large high-quality datasets. However, the lack of large-scale, publicly available AS datasets limits model development tailored to this patient cohort. Existing models are often trained on non-AS datasets, predominantly featuring high-risk PCa cases and single time-point MRI scans. However, a model trained on non-AS data can suffer from domain shift challenges when applied to AS data. This paper presents a novel approach using a dual-scan model to enhance prostate zone segmentation for patients in AS using longitudinal MRI scans of the same patient. Our dual-scan model effectively captures temporal changes in prostate anatomy by incorporating a cross-attention mechanism into a pre-trained 3D U-Net architecture. This improves prostate zone segmentation supporting lesion detection and facilitating more precise disease monitoring and treatment planning. Our method addresses the challenges of limited annotations in training data and domain shift, demonstrating significant performance gains over models using only single time-point approaches.