paulestin / retinal_oct

Retinal OCT

Retinal OCT

Problem Statement

Identifying human diseases from medical images. Using supervised and semi-supervised techniques. Retinal optical coherence tomography (OCT) is an imaging technique used to capture high-resolution cross sections of the retinas of living patients. Approximately 30 million OCT scans are performed each year, and the analysis and interpretation of these images takes up a significant amount of time (Swanson and Fujimoto, 2017).

Resources

• Introduction
• Publication
• Additional Information on the classification types
• Dataset

Pre-reqs

Complete Setup

Training

You can run the shortcut command tasks/train_retina_predictor.sh, which runs the following:

python training/run_experiment.py --save '{"dataset": "RetinaDataset", "model": "RetinaModel", "network": "resnetconv", "train_args": {"batch_size": 32}}'

Running sweeps (hyper param optimization using weights & biases).

You can parallely run many sweeps, below is one example

wandb sweep training/sweep_resnet_finetune.yaml
copy the sweepid from above
wandb agent {sweepid}

Findings

Supervised Learning

Network	Train Acc	Val Acc	Test Acc	Hyperparam Optimization
Resnet	84.6%	88.5%	93.2%	Sweep Config
Resnet FineTune	87.39%	91.26%	97.8%	Sweep Config

Self-Supervised Learning

Network	Contrastive Accuracy	Supervised Accuracy	Hyperparam Optimization
SIM CLR	97.4%	89.46%	Sweeps

Acknowledgements

• Self-Supervised Learning (SimClr)
• Full Stack Deeplearning
• Weights and Biases