Here is the github repo for paper EdgeMixup: Embarrassingly Simple Data Alteration to Improve Lyme Disease Lesion Segmentation and Diagnosis Fairness
There are two datasets: one for segmentation model and one for classification model.
For the segmentation dataset, we annotate skin images into three classes: background (black), skin (yellow), and lesion (blue). The lesion area contains three types of disease/lesions: Tinea Corporis (TC), Herpes Zoster (HZ), and Erythema Migrans (EM).
The classification dataset has 2,712 samples, and we annotate those skin images into four classes: No Disease (NO), TC, HZ, and EM.
python train_segmentation_model.py --regenerate-splits
python train_segmentation_model.py --train
python train_segmentation_model.py --test
Generate segmentation masks for the classification images calculate the image's ITA, and clean the data splits.
python classify_disease.py --generate-masks
Train a model: python classify_disease.py --train --model-type baseline
Evaluate a model:python classify_disease.py --test --model-type baseline
On Segmentation tasks: python train_segmentation_model.py --iterative
On Classification tasks:python classify_disease.py --train --edgemixup = True
@InProceedings{10.1007/978-3-031-43901-8_36,
author="Yuan, Haolin
and Aucott, John
and Hadzic, Armin
and Paul, William
and Villegas de Flores, Marcia
and Mathew, Philip
and Burlina, Philippe
and Cao, Yinzhi",
title="EdgeMixup: Embarrassingly Simple Data Alteration to Improve Lyme Disease Lesion Segmentation and Diagnosis Fairness",
booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023",
year="2023"
}