Note: 1) Please do not delete any existing files/folders under the folder "./PLDC_software/ ", such as image, and doc folders.
2) For image format input, make sure your input T2 image is a 3-channel image. Only the T2 image is allowed, as ADC and hDWI require metadata of T2 DICOM for prostate region registration.
3) For DICOM format input, T2 is necessary, while ADC and hDWI are optional.
4) We provide several samples (under the folder "./test_samples" ) from PROSTATEx to demonstrate the software functions, while the local cohort samples are not open-sourced here due to patient privacy and ethical issues. Our weights trained with PROSTATEx (source domain) and our local target dataset are also provided for demonstration. Our weights can be downloaded from the same Github repository. The total size of this software is around 6GB, including the executable software (~699MB) and model weights (~3.8G).
Prostate lesion assessment using your local cohort mpMRI
Prepare your multi-cohort mpMRI (T2, ADC, and hDWI) datasets for model training. mpMRI dicoms are required, as the metadata has to be extracted for prostate region segmentation on ADC and hDWI. The datasets include a source domain dataset (e.g., the public dataset
PROSTATEx ), and your mpMRI-based local cohort dataset (i.e., target domain). Our models (i.e. prostate segmentation model and CMD²A-Net) are trained with API Keras. You will train and test the models for PLDC in the following steps. We also offer our executable codes and files online available via the same Github repository, so as to allow any work extension or application by others.
Prostate region segmentation
To train the prostate segmentation model using T2 images, you can train your prostate segmentation model with the codes under ./maskrcnn_model. Apart from your local cohort datasets, the public dataset
I2CVB is also available online for training. The input shape of the model is set to 512 × 512 pixels. Adam optimizer is applied. During the training process, the model with the highest dice coefficient score is retained (i.e., weight1.h5). The prostate on T2 images can be segmented in the following code.