Deepfake Detection in Medical Images Using Mask RCNN

The rapid advancement of deep generative models has significantly increased the potential for manipulation of medical images, raising serious concerns regarding the reliability of diagnostic imaging and patient safety. This paper proposes a novel deepfake detection framework based on Mask R-CNN, leveraging instance segmentation to identify pixel-level alterations in brain MRI scans. The proposed approach integrates both classification and segmentation to categorize images into four classes: genuine tumor, no tumor, deepfake with tumor insertion, and deepfake with tumor removal. By utilizing instance mask outputs, the model effectively captures subtle structural inconsistencies that are often imperceptible through visual inspection. The framework is implemented as a web-based system using Next.js for the frontend and Python for the backend. The Mask R-CNN model is trained and validated on a curated dataset of 1,300 annotated brain MRI images. Experimental results demonstrate high accuracy in distinguishing authentic and manipulated scans, along with reliable segmentation of altered regions. The findings highlight the effectiveness of instance segmentation in detecting adversarial modifications and emphasize the importance of automated validation systems in preserving the integrity of medical imaging. Future work aims to enhance model robustness and expand the dataset for improved generalization.