We propose three novel VGG based architectures for prostate lesion classification (see Architectures module):
- CNN_VGG_SIMPLE with three subnetworks for mpMRI modalities
- CNN_VGG_MODALITIES with mixture of experts architecture (each modality is capable of making predictions)
- CNN_VGG_PIRADS with a-priori knowledge embedding in mixture of experts architecture
This project uses data from ProstateX1 competition.
Directory data/ProstateX contains trail dataset, a preselected cases from ProstateX1 Challenge. In order to perform analysis on bigger dataset - download the data from original sources and extract them to directory following the trail directory as an example.
Install Anaconda 4.3.1 (with Python 3.6.0) - https://repo.continuum.io/archive/ Execute setup.bat (or copy content to terminal on unix) Note that requirements.txt may not be complete - if so, please add missing requirements to the text file and make pull request
After installation, to run the tests:
- Run augment_data.py to augment and save locally the data
- Modify constants.py to match your experiment parameters
- Run test.py to train the model using CV and verbose logging