ISIC-2024 Challenge (Kaggle)

This repository is a "shallow fork" of a private competition repository containing my part of the work only on the team "fx-9860GII" for the ISIC 2024 Challenge hosted on Kaggle. The aim of the competition is to predict whether a skin lesion is benign or malignant, based on provided images and metadata of the lesion.

Results

The final selected solutions were based on gradient-boosting methods. The lesion images first passed into CV models to yield vision-only predictions. These predictions are concatenated to the feature-engineered metadata table as input for three different gradient boosting frameworks. Finally the predictions from the three frameworks are combined together to give a final score.

• Stratified 4-fold cross-validation trained BEiT -> gradient boosting: 0.17353 (public leaderboard (LB)), 0.16030 (private LB)
• ViT, ViTMAE, BEiT, EfficientNet-b7 -> gradient boosting: 0.17624 (public LB), 0.15708 (private LB)

Directories

• docs: Some notes.
• preprocess_data: Data preprocessing, basic dataset statistics and my attempts to combine data available on the ISIC Archive this year's new data.
• scripts: Scripts for remote GPU instance setup
• src: Most of the training and evaluation scripts.