DeepMedFuseX: Explainable DeepFake Cancer CT Scan Classification with Multi-Scale Attention and Transfer Learning

Abstract:
This paper introduces DeepMedFuseX, a novel deep learning framework developed for the accurate classification of cancer computed tomography (CT) scans as either authentic or synthetic (deepfake). The proposed approach leverages transfer learning by utilizing a pre-trained ResNet-50 model, which has been fine-tuned on a specialized CT scan dataset. This model is further enhanced with the Convolutional Block Attention Module (CBAM). This multiscale attention mechanism recalibrates feature maps to capture intricate details in medical images both channel-wise and spatially. Emphasis is placed on the interpretability of the model through the integration of the Gradient-weighted Class Activation Mapping (Grad-CAM) framework, underscoring the critical importance of explainability in AI models, particularly within the medical domain. The DeepMedFuseX framework demonstrates significant improvements in classification accuracy and robustness on challenging datasets, providing a powerful tool for medical practitioners in combating the threat posed by deepfake medical images.