Explanation supervision aims to improve deep learning models by integrating additional signals for generating model explanations. Challenges in supervising visual explanations in 3D data include altered spatial correlations, sparse annotations, and varying uncertainty. The proposed Dynamic Uncertainty-aware Explanation (DUE) framework addresses these challenges through diffusion-based 3D interpolation with uncertainty-aware guidance. Comprehensive experiments on real-world medical imaging datasets validate the effectiveness of the DUE framework in enhancing model predictability and explainability.
To Another Language
from source content
arxiv.org
Ключові висновки, отримані з
by Qilong Zhao,... о arxiv.org 03-19-2024
https://arxiv.org/pdf/2403.10831.pdfГлибші Запити