Deepfake MRI Detection: A ViViT-Based Landmark Feature Extraction Approach for Brain Tumor Scans

Deepfake manipulation of MRI brain tumor scans pose significant risks, including misdiagnosis, fraudulent research, and ethical concerns in medical imaging. To address this, we propose a Video Vision Transformer (ViViT)-based deepfake detection model enhanced with landmark-based feature extraction, designed to identify manipulated MRI scans with high accuracy. Our approach adapts ViViT to analyze sequential MRI slices, capturing subtle inconsistencies introduced by deepfake techniques, while landmark-based features focus on key tumor regions for improved precision. Additionally, Depthwise Separable Convolutions (DSC) and Convolutional Block Attention Module (CBAM) are incorporated to enhance feature learning and computational efficiency. The proposed model is evaluated on the BraTS 2021 dataset, achieving a Dice Similarity Coefficient (DSC) of 91.2%, Intersection over Union (IoU) of 85.7%, and an AUC-ROC of 98.4%, significantly outperforming CNN-based and GAN-forensic approaches. Based on the results, the proposed system exhibit higher accuracy and robustness.