Out-of-Distribution Detection for skin lesion images
This repository contains the code for the MICCAI 2022 paper "Out-of-Distribution Detection for Long-tailed and Fine-grained Skin Lesion Images" (https://arxiv.org/abs/2206.15186)
Note that the code is only for ISIC2019 dataset as our in-house dataset could not be publicly released.
There are four experiments to be trained as listed below.
- Baseline model with only Cross Entropy Loss without any Mixup strategies
python train.py --loss Softmax --mixup 0
- Model employing mixup strategies
python train.py --loss Softmax --mixup 1
- Model with only Prototype Loss
python train.py --loss GCPLoss --mixup 0
- Model with integration of Mixup strategies with the Prototype Loss
python train.py --loss GCPLoss --mixup 1
As can be noted the arguments of --loss
and --mixup
control the different experimental settings for the methods proposed.
For testing the above trained models, please follow the below commands.
- For the standard Cross Entropy loss trained models with / without Mixup strategies
python val.py --loss Softmax --checkpoint <checkpoint_path> --output-filename <output_filename.csv>
- For the Prototype Loss trained models with / without Mixup strategies
python val.py --loss GCPLoss --checkpoint <checkpoint_path> --output-filename <output_filename.csv>
Here the --checkpoint
corresponds to the checkpoint path where the trained model checkpoint has been saved and --output-filename
corresponds to the output filename where the testing result will be stored. The output filename should be given as a .csv file.