Implementation of two few-shot method (Baseline and Meta Baseline) for MedFMC
- The following setup has been tested on Python 3.9, Ubuntu 20.04.
- mmpretrain (recommended 1.0.0rc8): please refer to https://github.com/open-mmlab/mmpretrain for installation details.
- sklearn (recommended 1.2.2): pip install sklearn
- Run the script of 'run_extract_feats.sh' to extract the features via pretrained models (e.g. swin-base) of all the test images in each dataset (ChestDR, Endo, NeoJaudice, ColonPath, Retino). The extracted features would be stored as '.npy' format file in the sub-folder 'test_feats/' of the dataset path (e.g. ./data/ChestDR/test_feats/swin-base.npy).
- Run the script of 'run_fewshot_baseline.sh' to test the results of Baseline and Meta Baseline method using 1, 5, 10 shot samples per class under 10 iterations for each dataset.
- In our experiments, image list used for randomly picking support set of each dataset is saved as 'fewshot-pool.txt', meanwhile image list consisted of the remaining testing images as 'test.txt'. These two image list files can also be found in data folder (e.g. ./data/ChestDR/test.txt).
- When you start to test this baseline, please place all the preprocessed images of each dataset into the sub-folder 'images/' beforehand (e.g. ./data/ChestDR/images/).
Wang, D., Wang, X., Wang, L. et al. A Real-world Dataset and Benchmark For Foundation Model Adaptation in Medical Image Classification. Sci Data 10, 574 (2023). https://doi.org/10.1038/s41597-023-02460-0