jacob-thrasher / AL-OCTA

OCTA500-classification

This repository contains supplementary code for linkEnhancing Retinal Disease Classification from OCTA Images via Active Learning Techniques, published in Data Engineering in Medical Imaging (DEMI) workshop at MICCAI 2024. In this work, we analyze active learning as a data subset technique to improve classification performance on OCTA images.

Figure: Schematic diagram of active learning pipeline, where $`\mathcal{D}_{train, val}`$ are the train and validation set, respectively, X is an image, s is the uncertainty score, and k is the number of images to move into the train set after each active learning iteration.

Dataset

We utilize the OCTA500 dataset, which contains images of healthy, age-related macular degeneration (AMD), chorodial neovascularization (CNV), and diabetic retinopathy (DR) retinas. More details can be found link here

Figure: Comparison of OCT and OCTA data for Normal, CNV, DR, and AMD eyes.

Classification results

Training Method	Acc	F1
Unbalanced	.5139	.4864
Inverse Frequency Class Weighting	.4956	.4571
Random Undersampling	.4482	.3334
Oversampling (AutoAugment)	.4178	.4136
Oversampling (AugMix)	.4647	.4503
Least Confident Sampling	.7313	.6285
Entropy Sampling	.7188	.6187
Margin Sampling	.7282	.6262
Ratio Sampling	.7688	.7116