HINTS
From black box to intelligible machine learning for the accurate diagnosis of medical images
Description
In this project we investigate how “hints” - additional annotations of the visual content of the image - can help medical image classification. We have shown that in skin lesion classification, annotations of high-level properties such as asymmetry of the lesion, can be used in multi-task learning to improve the robustness of the algorithm. Additionally, such annotations may help for the algorithms to be more explainable.
People
Ralf Raumanns (principal investigator), Veronika Cheplygina.
Funding
NWO (Dutch Research Council) Lerarenbeurs
References
-
arXiv preprint arXiv:2407.17543, 2024 -
MELBA (Machine Learning for Biomedical Imaging), 2021
