Domain adaptation model training for mpMRI-based PLDC

For DA model (i.e., CMD²A-Net) training, the prostate regions from both the source and target domains are scaled to 224 × 224 pixels. Random rotation is applied for data augmentation. Adam optimizer is chosen. We initialized our model with the pre-trained CM-Net, in order to facilitate its convergence. To be specific, we trained both the coarse segmentation module and classifier of CM-Net first using codes under ./joint_model,  with the combined samples from both domains,  the best models are saved ( i.e., T2: weight2_1.h5, ADC: weight3_1.h5, hDWI: weight4_1.h5 ). Then, we optimized the total loss of CMD²A-Net with the source and target samples. By co-training all the modules, the models with the highest accuracy (i.e., T2: weight2_2.h5, ADC: weight3_2.h5, hDWI: weight4_2.h5) are saved for testing. The PLDC testing in the target samples from local cohort can be achieved by the following codes: