A pseudo-3D multi attention mechanism for prostate zonal segmentation

Accurate segmentation of small regions in medical imaging is vital for monitoring clinical pathology and treatment efficacy. Traditional 2D attention and transformer mechanisms, while effective at capturing dependencies, struggle with volumetric data. Conversely, 3D neural networks excel at capturing spatial dependencies across depth, height, and width but are computationally expensive. To address these challenges, we propose a pseudo-3D approach using the multi-attentive slice constrained (MASC) network architecture, designed to enhance prostate peripheral (PZ) and transitional (TZ) zone segmentations. The MASC network integrates pseudo-3D attention mechanisms to efficiently capture contextual dependencies within and between slices, balancing computational efficiency with the ability to incorporate essential volumetric context. Our approach involves a novel MASC module combined with a U-Net backbone. The module utilizes a multi attention ladder mechanism to process feature tensors, generating attention maps that are concatenated to form the final attention feature map. This map is used in a multi-scale encoder setup to capture extended-range dependencies without the overhead of full 3D networks. The MASC incorporates slice interaction to simulate 3D attention in a computationally efficient manner. Evaluated on a dataset of 44 patients with high-resolution T2W MRI images, our model demonstrated superior segmentation performance compared to 2D and 3D models, particularly in distinguishing between PZ and TZ zones. The MASC approach effectively simulates volumetric information, improving accuracy while maintaining computational efficiency. A segmentation accuracy of 0.85 and 0.65 was obtained for the transitional and peripheral zones respectively. Future work will explore integrating MASC with various backbones to further enhance segmentation performance.