Introduction

Repository for "Reliable Multi-View Learning with Conformal Prediction for Aortic Stenosis Classification in Echocardiography" Includes code to train ResNet-18 and R(2+1)D models with the RT4U algorithm using the Pytorch Lightning framework Includes code for conformal prediction algorithm and interpretation of results in jupyter notebooks

Installation

Docker

• Initialize a docker container running pytorch at https://hub.docker.com/r/pytorch/pytorch/tags. The majority of the code is ran using pytorch:1.12.1-cuda11.3-cudnn8-runtime
• Pytorch version may be changed depending on which CUDA version is on your local device
• refer to docker_setup for example
• Run setup.sh within the container to install relevant libraries

Conda

• Configure a conda environment with a gpu-compatible pytorch version
• Install the list of files in setup.sh

Preparing data

Data folders are assumed to be stored in the /data volume. To change this behavior, modify the src/*.yaml files

• The folder should be organized as follows for CIFAR within /data/cifar-10-batches-py/ (installing via torchvision.datasets.CIFAR10 is fine)
  • data_batch_[1..5]
  • metadata
• The folder should be organized as follows for TMED within /data/TMED/approved_users_only/
  • DEV[165/479/56] folders
  • dataset folders
  • csv files
• The folder should be organized as follows for the private AS dataset within /data/Aortic_Stenosis/as_tom/round2/
  • plax folder
  • psax folder
  • sheets folder
  • annotations-all.csv

Usage

YAML files

• Configure settings and hyperparameters in the config_x.yaml files
• For training without initial checkpoint, set test_only to false and "ckpt_path" to null
• For training with checkpoint, set "ckpt_path" to the path to a checkpoint file
• For testing, set ckpt_path and then set test_only to true
• Files created during the run are stored under a newly created folder, the folder can be configured in the logger section of the YAML

Running

• Refer to main_multi_round.py for the training flow
• We use Hydra to integrate the yaml settings into the training process, refer to training_runs_x.sh for examples

Checklist before your first run

• Did you install relevant libraries in setup.py?
• Did you put the data in the relevant folder?
• Did you read the .yaml file and edit relevant fields?
• Do you understand where the output files from your training will go?