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

2026-08-05

AI Summary

This work addresses the limited generalizability and poor interpretability of existing deepfake detection methods in medical imaging, which struggle to handle CT scans manipulated by unseen generative models. To overcome these challenges, the authors propose HexMIL, a multiple instance learning framework that requires only volume-level binary labels—eliminating the need for pixel-wise annotations—to simultaneously detect and localize forgeries. Its key innovation lies in a hierarchical attention mechanism: by integrating a two-level patch-and-slice partitioning strategy with dedicated gated attention modules, HexMIL produces full-resolution 3D attention maps whose weights directly inform classification decisions, yielding spatially faithful and structurally coherent attribution. Experiments demonstrate that under cross-generator generalization settings, HexMIL outperforms baseline methods by 9.1% in AUC and 9.4% in F1 score, while achieving state-of-the-art performance in localization as measured by mean IoU and Pointing Game accuracy.

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.

Problem

• medical deepfakes
• manipulation detection
• generalization
• interpretability
• CT volumes

Innovation

• Multiple Instance Learning
• Ante-hoc Explainability
• Hierarchical Attention
• Medical Deepfake Detection
• Cross-generator Generalization