HexMIL: Hierarchical Attention MIL for Ante-Hoc Explainable Detection of AI-Manipulated CT Volumes

HexMIL is a mask-free detector for AI-manipulated CT volumes trained with only binary volume-level labels. It decomposes each volume into a hierarchy of patches and slices, then combines independent Gated Attention modules into a full-resolution 3D attention volume. Because these weights directly participate in the forward classification computation, the method claims structurally faithful, ante-hoc localization rather than post-hoc explanations such as Grad-CAM. On M3DSynth and CT-GAN, cross-generator evaluation reportedly improves out-of-domain performance over all baselines by 9.1 AUC and 9.4 F1, while achieving the best average IoU and Pointing Game localization scores.