Agentic Mixture-of-Experts Forensics for Detecting and Localizing Deepfake Manipulations in Volumetric Lung Computed Tomography

Deepfake manipulation of volumetric CT scans threatens diagnostic reliability and radiological security, as modern generative models can alter tumor regions with high anatomical realism. This study introduces a hybrid forensic framework combining 2D slice classification, 3D cube-based volumetric modeling, coordinate-aware tamper localization, and a Mixture-of-Experts (MoE) architecture to detect and interpret CT-GAN-based manipulations. Experiments on a tampered lung CT dataset show that the 2D model achieves the strongest and most stable detection performance (0.862 test accuracy), while the 3D cube model reaches comparable accuracy but offers limited volumetric advantage under the small-sample regime. The coordinate-regression network localizes manipulated tumor centers with a mean error of 205 px, indicating that subtle spatial irregularities remain learnable despite low visual contrast. The two-expert MoE classifier also attains 0.828 accuracy but exhibits pronounced specialization, with the gating network routing most slices to a single expert during validation and testing, an emergent agentic behavior reflecting expert-specific forensic cue learning. Together, these findings demonstrate that slice-level reasoning, coarse spatial localization, and adaptive expert routing provide complementary forensic signals for deepfake CT detection. The proposed framework lays the groundwork for multi-stage, tamper-aware radiology systems that integrate interpretability, robustness, and security within clinical imaging workflows.