Explainable Feature-Injected Diffusion Model for Medical Image Translation

The integration of computed tomography (CT) and magnetic resonance (MR) imaging is crucial for accurate medical diagnosis and treatment planning. However, translating images between CT and MR remains challenging due to significant differences in imaging modalities. To address this problem, we propose an Explainable Feature-Injected Diffusion Model (EIDM) for unsupervised CT-to-MR image translation. EIDM comprises a feature synthesis module and a diffusion-based latent space learning framework. This model captures frequency representations of the original CT images using the Fast Fourier Transform and applies high-pass filters to restore anatomical structures lost during diffusion. It also integrates weighted heatmaps generated by explainable AI models and utilizes a cross-attention mechanism to achieve unbiased image synthesis. We quantitatively evaluated EIDM and recent approaches using four metrics for comparison. Experimental results demonstrate that EIDM outperforms latest Generative Adversarial Networks (GANs) and diffusion models, generating realistic MR images that preserve anatomical integrity, as evidenced by enhanced scores across evaluation metrics. This work highlights the effectiveness of jointly learning explainable features and contour regions in achieving the goal of translating CT to MR images.