Deep learning algorithm for predicting Gleason grade group based on prostate MRI image set

Accurate Gleason Grade Group (GGG) prediction in prostate cancer is crucial for prognosis and treatment decisions. Traditional biopsy, however, suffers from invasiveness, sampling error, and inconsistent pathological scoring. Thus, non-invasive, stable, and accurate imaging tools are urgently needed. In this study, we propose a Prostate Dual-branch Hybrid domain prediction Network (PDHD-Net) based on Multi-parametric MRI (mp-MRI) data to enable fine-grained Gleason Grade Group (GGG 1-5) stratification. we propose an innovative frequency-band-shunting feature-enhancement strategy. Specifically, Leveraging high-frequency information to enhance the subtle textural differences between lesion areas and surrounding tissues, while utilizing low-frequency information to accentuate overall structural and morphological patterns, through combining deep convolutional networks' sensitivity to local fine-grained features with self-attention mechanisms' capability to capture global semantic relationships achieves more accurate identification of lesion regions and Gleason Grade Group (GGG) prediction in MRI images. The FROC analysis of the multi-center dataset (including the public ProstateX-2 and private datasets) indicates that: PDHD-Net achieved sensitivities of 96.7% and 88.4% for high-risk (GGG=5) and clinically significant lesions, respectively, with an average of one false positive per patient. Compared with previous studies that uniformly grouped high-risk lesions into one group, this study achieved a fine distinction between GGG=4 and GGG=5. Furthermore, for the GS classification, evaluated by the receiver operating characteristics (ROC) analysis, PDHD-Net received the area under the curve of 0.82 for discriminating clinically significant PCa (GGG≥2) from GGG=1. Notably, the Dice of which distinguished Gleason's five classifications reached 0.80. The source code is available at https://github.com/NGYLK/PDHD-Net