Deep neural networks can achieve strong predictive performance in medical imaging while relying on regions that have nothing to do with the finding being predicted. A new arXiv paper argues that this makes models less trustworthy, since their decisions may be driven by irrelevant structures or artifacts rather than the actual pathology. Explanation-guided learning is one proposed remedy, but the abstract notes that multi-label classification makes the problem more complicated.
The paper, titled "Learning from the Right Regions: Class-Conditioned Explanation Guidance for Medical Imaging," suggests conditioning the explanation guidance on the class being predicted. This is meant to encourage the model to use regions relevant to each specific finding, rather than applying a generic explanation constraint across all labels.
The source is a single arXiv abstract, so the write-up is limited to the proposed framing. No experimental results or implementation details are included in the available text, and there are no other sources to compare against.