Project Blog Post here.
- Authors are Austin Geary, Tim Strebel, and Will Dougall.
- This project was conducted as the Captsone Project for the Masters in Applied Data Science (MADS) program in the School of Information at University of Michigan.
- We attempted to create a CNN model that would perform at a level sufficient for Radiology departments in clinical settings to use to determine the presence or absence of pneumonia in patient lungs.
- The reason for this undertaking is that skilled radiologists are in short supply and are often over-worked. AI can relieve the workload of human radiologists as well as accelerate the process of providing test results to physicians.
- Modular collection of scripts for model training. Can use DenseNet121, ResNet18, and AlexNet as well as two different datasets called RSNA and CX14.
- Notebooks for producing visualizations of patient metadata, as well as model evaluation. A technique called grad-cam was used to visually highlight areas of activation by the model for positive cases.
- Utilized PyTorch lightning to abstract away parallelization and model training loops to streamline code.
Environment requirements are listed in a file called requirements.txt located in the root of the repository. Datasources are listed in a file called img-data-source-readme.txt located in the data folder.
For training the models, open up a command line, navigate to the src directory containing train.py, and then execute the script and pass in arguments found in the args.py file.
Example:
python3 train.py --model densenet --epochs 5 --targets_path ../../data/rsna-targets.csv --image_dir ../../data/chest-xray-14/images --freeze_features All --init_learning_rate 3e-3
For running notebooks, simply launch a Jupyter Notebook server session and navigate to the notebooks directory.
Project is: no longer being worked on. The reason for this is our semester came to a close, but there are many other paths we could take to build upon the work done thus far.
- This project was based on this Kaggle competition.
- Many thanks to the MADS Staff at the University of Michigan
- Many thanks to Dr. Amilcare Gentili and Dr. Michael J. Kim from the VA Healthcare system for agreeing to be interviewed for our project.
Created by @Tstrebe2, @austingeary and @Zbandit98