ECGenius

Steps to run

1. Run ./setup_env.sh
2. Make sure all dependencies in requirements.txt are installed

File descriptions

dicom_loader.py: Implements a dataloader over torch.utils.data.Dataset for converting Echo-Cardiogram videos (*.dcm files) to torch tensors, passing them through preprocessing

preprocessing.py: Implements transforms over tensors read obtained from dcm files for remaining superfluous information

requirements.txt: Packages needed for running ECGenius

results_generator.py: Script version of python notebooks for running models for a longer duration

runner.py: Implements generic Trainer class to abstract out training and evaluation process

setup_env.sh: Creates symlinks for data and validation directories, creates environment.json for various system-specific parameters that would otherwise need to be hardcoded

checkpoints/: Saves model weights and loss history plots

docs/: Code for GitHub page for the project

models/: Contains code for all the models tested until now

• models/auxiliary/: Models related to the 2nd dataset (Heart failure clinical records) and insights from the RVENet dataset
  • models/auxiliary/RandomForestRegressor.py: Identify importance of age, sex, patient group and heart rate for prediction of RVEF
  • models/auxiliary/text_preprocessing.py: Preprocesses and removed outliers from heart failure clinical records dataset
  • models/auxiliary/heuristic_concatenation.py: Concatenates cardiovascular and RVENet dataset by bucketizing age groups in 0-100
  • models/auxiliary/merge_datasets.py: Tries pandas merge + iterative imputation to concatenate cardiovascular and RVENet dataset
  • models/auxiliary/merge_dataset_knn.py: Tries k-nearest neighbors to concatenate cardiovascular and RVENet dataset
  • models/auxiliary/*.csv: Corresponding merged dataset outputs
  • models/auxiliary/get_summary_llm.py: Get LLM summary for the patient personal merged from 2 datasets
• models/rvenet/: Deep neural networks implemented for the RVENet dataset for predicting RVEF (Right ventricle ejection fraction) from echo-cardiogram videos
  • models/rvenet/all_models.py: Saves union of all models implemented so that it doesn't need to be hardcoded
  • models/rvenet/CardiacCycleRNN.py: Resnet18 + LSTM + Linear layer + Feature Augmentation
  • models/rvenet/CardiacCycleTransformer.py: Resnet18 + Transformer Encoder + Linear layer
  • models/rvenet/ResNetLSTM.py: Resnet18 + LSTM + Linear layer with different loss function
  • models/rvenet/ResNextTemporal.py: Temporal CNN with ResNext as backbone

notebooks/: Cleaned up versions of notebooks used during experimentation in the project notebooks/feature_augmentation_benefits.ipynb: Notebook to experiment if feature augmentation helps in CariacCycleRNN

setup/: Directory storing environment.json and containing python script for generating it