GRASP-Net: Grouped Residual Convolution U-Net With Attention Mechanism and Atrous Spatial Pyramid Pooling for Prostate Zone Segmentation Using MR Images

Abstract

Prostate cancer is a prevalent disease in men, especially among the elderly, and magnetic resonance imaging is the leading acquisition method for the diagnosis and evaluation of the prostate. Accurate segmentation of the prostate, particularly the transition zone and peripheral zone, is crucial for early detection and effective treatment planning. This work introduces GRASP-Net as an innovative deep learning-based model to improve prostate MRI zonal segmentation accuracy. GRASP-Net integrates grouped residual convolutional modules, attention mechanisms, convolutional block attention module, and atrous spatial pyramid pooling blocks to enhance feature extraction and boundary segmentation. The model has been evaluated on the Medical Segmentation Decathlon Task 05 Prostate dataset, comparing its performance against other well-known models. Overall, the GRASP-Net model achieved higher segmentation results with a dice similarity coefficient of 0.928 for the transition zone and 0.864 for the peripheral zone, surpassing previous state-of-the-art results. Additionally, the model exhibits significant performance on 95 percentile Hausdorff Distance, Average Surface Distance, and Sensitivity values and proving its accuracy in anatomical prostate structure localization. These advancements emphasize the promising prospect of the GRASP-Net model to advance prostate cancer diagnosis and treatment, presenting an effective tool for clinical usage.