Predicting Pulmonary Edema Using Deep Learning and Image Segmentation
Team Members: David Davila-Garcia, Marco Morocho, Yash Potdar
Note: All data was deidentified but is not publicly available
├── README.md <- The top-level README for developers using this project.
├── Final_Report.pdf
├── Final_Poster.pdf
├── models <- Contains the outputs from trained models: Losses and Test Set Predictions
│ ├── Losses <- Training and Validation MAE Losses by Epoch.
│ ├── Test Set Preds <- NT-proBNP predictions on test set using the best model (minimize MAE valid loss).
├── 1 - Preprocessing.ipynb <- Cleaning the original x-rays + clinical data, excluding rows with missing data/no image available
├── 2 - Transfer Learning Training & Evaluation.ipynb <- (Not used in project) Provided by UCSD AIDA Lab, shows training of U-Net segmentation model.
├── 3 - Predicting Unannotated.ipynb <- Used the U-Net segmentation model from the UCSD AIDA Lab to create binary masks (lungs, heart, clavicles, spinal column) for each radiograph in our dataset. Saved the segmentations to an hdf5 file.
├── 4 - Creating Masks.ipynb <- Uses the binary masks created in '3 - Predicting Unannotated.ipynb' to produce the segmentation inputs for our model
├── 5 - CNN Models.ipynb <- Contains all code for training and testing models.
├── model.py <- Contains modified ResNet152 architectures, extends the Pytorch ResNet152 implementation
├── train.py <- Contains model training and testing functions; different inputs called for different architectures
Acknowledgements: Thank you to our incredible mentor Albert Hsiao, MD, PhD for his guidance, and Amin Mahmoodi for providing the U-Net segmentation network.