Abstract
Clinical applicability of automated decision support systems depends on arobust, well-understood classification interpretation. Artificial neuralnetworks while achieving class-leading scores fall short in this regard.Therefore, numerous approaches have been proposed that map a salient region ofan image to a diagnostic classification. Utilizing heuristic methodology, likeblurring and noise, they tend to produce diffuse, sometimes misleading results,hindering their general adoption. In this work we overcome these issues bypresenting a model agnostic saliency mapping framework tailored to medicalimaging. We replace heuristic techniques with a strong neighborhood conditionedinpainting approach, which avoids anatomically implausible artefacts. Weformulate saliency attribution as a map-quality optimization task, enforcingconstrained and focused attributions. Experiments on public mammography datashow quantitatively and qualitatively more precise localization and clearerconveying results than existing state-of-the-art methods.