MedFakeDet: multi-organ medical deepfake detection and inpainting localization system via effective deep learning models and density maps

The ability of deepfake technologies to produce highly realistic images has become a significant security concern for healthcare systems and insurance auditing processes, either through the generation of fake medical images or the realistic manipulation of original ones. This study employs effective deep learning methods to detect medical images generated by deepfake technologies and to localize inpainting-based manipulations. Two different datasets were constructed for this purpose: the MedFake dataset, which contains brain, lung, kidney, and chest images generated using multiple synthesis approaches; and the MedFakeInpaint dataset, which includes localized inpainting manipulations applied to brain images. For deepfake synthesis detection, an 8-class ResNet-based model was trained to jointly classify the organ type of a medical image and its real or fake status. For manipulation localization, a U-Net-based model was trained to predict density maps corresponding to manipulated regions within the image. Bounding boxes were subsequently derived from the predicted density maps to explicitly represent the localized manipulated regions. The performance of the proposed approach was evaluated using F1-score, and IoU-based localization metrics under strict training, validation, and test data splits with varying threshold values. The developed models were integrated into a backend system using FastAPI, and a lightweight web application was implemented to connect the provided endpoints with end users. MedFakeDet is released as a reproducible software package, including pretrained models and inference scripts. When the developed models were evaluated under different ablation configurations and threshold values, it was observed that in some configurations, the F1 score reached levels of up to 99%.