Training-free Test-time Improvement for Explainable Medical Image Classification

Hangzhou He#, Jiachen Tang#, Lei Zhu, Kaiwen Li, Yanye Lu $^*$

At present, the code is somewhat disorganized and we are cleaning up the code. We are also in the process of building another repository for CBMs training and user intervention, and then integrate the strategies proposed in this paper into that repository.

Datasets

Skin images: Fitzpatrick17k, DDI

WBC images: PBC, RabbinWBC, Scirep.

(The Scirep dataset needs to be obtained in communication with the paper authors at Scirep)

File Explanation

train_test.py: without introducting the confusion concept identification strategy, most of the code is the same as align_concept_cbm, except we generate a validation set for the fitz17k dataset to select the best val model for testing.

test_try1_skincon_impr.py, test_try1_pbc_impr.py: confusion concept identification and activation manipulation.

Acknowledgements

Our project and code are heavily based on align_concept_cbm, thanks for the great work and help from the authors.