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

Authors: Orazio Pontorno, Luca Guarnera, Zahid Akhtar, Sebastiano Battiato

Published: 2026-08-05 17:40:29+00:00

Comment: Accepted at ACM Multimedia 2026 (MM '26)

Journal Ref: Proceedings of the 34th ACM International Conference on Multimedia (MM '26), November 10--14, 2026, Rio de Janeiro, Brazil

AI Summary

HexMIL is a mask-free medical deepfake detector for CT volumes that addresses both poor generalization to unseen generative architectures and lack of interpretability. It uses a hierarchical attention-based multiple instance learning approach to decompose CT volumes and aggregate features, generating a full-resolution 3D attention volume for localizing manipulated regions without pixel-level supervision. HexMIL demonstrates superior performance in out-of-domain classification and localization compared to existing methods.

Abstract

The emergence of medical deepfakes, i.e., medical images manipulated by deep generative models, poses a significant threat to clinical workflows. However, existing detectors suffer from two critical limitations: poor generalization to unseen generative architectures for manipulation detection and lack of interpretability. In this context, we present HexMIL (Hierarchical EXplainable Multiple Instance Learning), a mask-free medical deepfake detector that simultaneously addresses both limitations using only binary volume-level supervision. HexMIL decomposes each CT volume into a two-level hierarchy of patches and slices, aggregated via independent Gated Attention modules whose weights are directly combined into a full-resolution 3D attention volume that localizes the manipulated sub-region without any pixel-level annotation. Unlike post-hoc methods such as Grad-CAM, HexMIL's attention weights constitute the exact forward computation driving the classification decision, providing ante-hoc and structurally faithful spatial attribution. We evaluate HexMIL on M3DSynth and CT-GAN datasets under a rigorous cross-generator generalization protocol, training on a single generative architecture and testing on unseen ones. HexMIL outperforms all baselines by $+9.1$ AUC and $+9.4$ F1 in out-of-domain classification, and achieves the best average IoU and Pointing Game score in localization. Project page: opontorno.github.io/hexmil.

Key findings

HexMIL significantly outperforms all baselines in out-of-domain classification with +9.1 AUC and +9.4 F1, demonstrating robust cross-generator generalization. It also achieves the best average IoU and Pointing Game score in localization, providing intrinsic and reliable spatial explanations compared to post-hoc methods. The hierarchical MIL architecture and gated attention mechanism are crucial for its superior performance and interpretability.

Approach

HexMIL decomposes each CT volume into a two-level hierarchy of patches and slices. It uses independent Gated Attention modules at both levels, whose weights are combined into a 3D attention volume to localize manipulated sub-regions. The attention weights provide ante-hoc and structurally faithful spatial attribution, driving the classification decision directly.

Datasets

M3DSynth, CT-GAN

Model(s)

ResNet-50 (as patch encoder), ResNet50-ABMIL, ViT-ABMIL (as baselines), R3D-18, DenseNet121-3D, ViViT (as baselines)

Author countries

Italy, USA

← Previous