Efficient Brain Tumor MRI Manipulation Detection Using a ShuffleNetV2-Based 3-Block Combination Model

This study identifies that texture-based features are crucial for detecting GAN-generated manipulated brain tumor MRI images, and it evaluates the efficiency of a lightweight detection model based on ShuffleNetV2. Two independent public MRI datasets were used to construct a multi-source environment, and paired Real–Fake images (B/B′, C/C′) were generated using StyleGAN2. Seven techniques—LBP, CLBP, Sobel, CBAM, PatchWise-SE, DropBlock, and TopK—were combined into 65 configurations with up to three modules. Experimental results showed that 3-Block combinations consistently outperformed single or 2-Block models in both accuracy and F1-score. The LBP + DropBlock + CBAM configuration achieved the best detection performance and generalization capability across domains. Most combinations maintained an Inference Time of 7-12 seconds. demonstrating the feasibility of real-time or near-real-time clinical application. These findings confirm that combining texture-based features with lightweight attention and regularization modules is effective for detecting manipulated medical images and highlight the potential of ShuffleNetV2-based lightweight models for practical deployment in clinical environments.