Explainable Artificial Intelligence (XAI) for 3D Medical Imaging: A Scoping Review of Volumetric Attribution Methods
TL;DR
Findings identify a gap between volumetric model inference and explanation presentation and motivate improved volumetric visualization and more consistent evaluation and identify the most frequently reported method family.
Abstract
Explainable artificial intelligence (XAI) methods developed for two-dimensional images may not preserve spatial context when applied to volumetric medical imaging. This scoping review map attribution methods for 3D medical images and examines how dimensionality is preserved across model inference, attribution generation, and visualization. Following PRISMA-ScR guidance, we searched IEEE Xplore, PubMed, ACM Digital Library, Semantic Scholar, and arXiv for studies published from January 2018 to May 2026. We mapped 32 selected studies by attribution method, imaging modality, clinical task, evaluation approach, and dimensionality paradigm. Grad-CAM and its variants were the most frequently reported method family. Within the selected corpus, two studies were classified as P2 boundary cases, 29 as P3, and one as P4. Most studies using volumetric inference presented explanations through two-dimensional views, and quantitative evaluation of attribution quality was uncommon. These findings identify a gap between volumetric model inference and explanation presentation and motivate improved volumetric visualization and more consistent evaluation.