SemanticIR: A Semantic-Aware Transformer for High-Quality Medical Image Restoration

Abstract

Background: Medical image restoration plays an important role in improving the quality and reliability of clinical imaging, including low-dose CT denoising, MRI super-resolution, and PET synthesis. Although CNN- and Transformer-based methods have achieved promising results, many existing approaches perform feature aggregation based mainly on fixed spatial windows or global pixel-level attention, which may disrupt coherent anatomical regions or introduce interference from semantically unrelated structures. To address this limitation, this study proposes SemanticIR, a semantic-aware Transformer framework for high-quality medical image restoration. Methods: SemanticIR incorporates semantic priors generated by the Segment Anything Model to guide restoration within and across anatomical regions. The proposed framework includes Region Attention for intra-region feature compensation, a Semantic-Aware Mixture-of-Experts module for region-specific feature transformation, and Cross-Region Attention for controlled inter-region information exchange. The model was evaluated on three representative medical image restoration tasks: CT denoising, MRI super-resolution, and PET synthesis. Performance was compared with multiple CNN- and Transformer-based baseline methods using PSNR, SSIM, VIF, and RMSE. Results: Experimental results demonstrate that SemanticIR consistently outperforms competing methods across all three restoration tasks. For CT denoising, SemanticIR achieved 46.97 dB PSNR, 0.9623 SSIM, 0.8282 VIF, and 1.7246 RMSE. For MRI super-resolution, it obtained 45.26 dB PSNR, 0.9783 SSIM, 0.8862 VIF, and 1.9691 RMSE. For PET synthesis, it achieved 40.99 dB PSNR, 0.9712 SSIM, 0.5933 VIF, and 2.7512 RMSE. Visual comparisons further show that SemanticIR restores clearer anatomical structures, sharper boundaries, and more faithful intensity distributions than baseline methods. Ablation studies confirm the effectiveness of Region Attention, Cross-Region Attention, and the Semantic-Aware Mixture-of-Experts module. Conclusions: SemanticIR provides an effective semantic-guided solution for medical image restoration by combining intra-region refinement, region-specific processing, and anatomically meaningful inter-region interaction. The proposed method improves restoration accuracy, structural consistency, and visual quality across multiple imaging tasks, showing strong potential for enhancing medical image quality in clinical applications.