This repository contains the code to reproduce the experiments in the paper "Revisiting Skin Tone Fairness in Dermatological Lesion Classification", accepted at the FAIMI workshop 2023 @ MICCAI 2023.
The dataset is the ISIC18 dataset (Task 3). It can be downloaded here. The images should be in the same folder as the metadata.csv.
The skin tone estimations based on the Individual Topology Angle (ITA) with the Deep Learning-based Healthy Skin Segmentation (DLHSS), the ITA labels provided by the authors of Kinyanjui et al. are the estimated_ita in the metadata.csv (data/.../metadata.csv).
For the skin tone estimation with Random Patch approaches, with arctan (RP) and with arctan2 (RP2), the code in the BevanCorrection notebook has been slightly adapted from Bevan et Atapour-Abarghouei.
The results and their visualisations can be reproduced with the predictions folder and with the Results_Plots notebook. Results regarding the comparison of skin tone estimates, prior to lesion classification, have been created with Tableau and can be reproduced with the Bevan_corrected.csv (RP,RP2) and with the ITA labels within the metadata.csv.
The experiments including the model training and grid search optimization of the MobileNetV2 are in the python files "*Baseline*.py" and "*DataShift*.py".
Also the optimization of the baseline classifier has been included in the 00Baseline02GridOptimization.py