6/7/2018 Steps to try to replicate results:
- Download PROSTATEx-v1-doinlp.jnlp file from ProstateX website :https://wiki.cancerimagingarchive.net/plugins/servlet/mobile?contentId=23691656#content/view/23691656
- Depending on your operating system, find a way to open and run the file to the get the dicom image data: https://fileinfo.com/extension/jnlp. This is a lot of data (1.5 GB unaugmented)
- Navigate to PROSTATEX directory and run the dicom_reader.py script. This will convert all of the T2-weighted and ADC images from dicom files to png files. Then run the generate_datasets.py or generate_datasets_b.py (with data augmentation).
Note on training these GANs: The group used the Sherlock Cluster at Stanford University to train these models on GPU nodes. For specs on the hardware, please see: https://www.sherlock.stanford.edu/docs/overview/specs/. Depending on the hyperparameters/job scheduling, these can take anyhwere from 6 hours to 2 days to run on this hardware.
- To run the DualGAN code simply copy over the datasets directory and its contents to the DualGAN/datasets directory. Alternatively, one can append the initial file path. (This will be cleaned up in a later version). Then follow the instructions as specified in the local directories README.md
- To run the CycleGAN code, do the same as specified above.
- Upon finishing the runs, run the distance.py and neighbor.py scripts to generate the quantitative and qualitative results.