2026-10-06T07:58:10Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34020477 2026-09-29T02:10:07Z category_35618 category_35825 category_35867 category_35549 category_35678 portal_1005 item_type_6 month_year_09_2026

Information Filter-based Diffusion Transformer for medical image super-resolution H Lu (5833871) X Wei (8427948) D Kong (13411254) J Xu (6794354) Frank Jiang (13070274) Computer vision and multimedia computation Information systems Library and information studies Information and computing sciences Data management and data science Science & Technology Technology Computer Science, Information Systems Information Science & Library Science Computer Science Medical image Super-resolution Diffusion model NETWORK Biomedical Imaging Bioengineering Lung Medical image super-resolution (SR) aims to recover fine anatomical details from low-resolution (LR) scans, yet obtaining high-resolution images is often challenging due to clinical limitations, such as acquisition time and radiation exposure. To address the limitation of conventional diffusion models with uniform noise perturbation in preserving sparse but clinically important structures, we propose an Information Filter-based medical image SR Network (IFNet), which incorporates region-aware guidance into the diffusion process. Specifically, we design a sparsity extractor to identify information-rich regions and generate a sparse structural mask, which is subsequently utilized to guide uncertainty-aware noise modulation and the denoising process. Furthermore, we develop an Information Filter-based Diffusion Transformer (IF-DiT) to enhance feature representation by adaptively emphasizing diagnostically relevant regions, thereby improving both reconstruction accuracy and computational efficiency. Extensive experiments are conducted on three public medical datasets, including IXI, BraTS2017, and Coltea-Lung-CT-100 W, under both (Formula presented) and (Formula presented) SR settings. The experimental results demonstrate that IFNet consistently achieves superior performance compared with recent state-of-the-art methods. Specifically, IFNet obtains PSNR/SSIM values of 50.27/0.9960 on IXI at (Formula presented) SR, 39.59/0.9860 on BraTS2017 at (Formula presented) SR, and 58.64/0.9992 on Coltea-Lung-CT-100 W at (Formula presented) SR. Compared with the strongest competing approaches, IFNet improves PSNR by up to 6.88 dB on IXI at (Formula presented) SR and 2.44 dB on BraTS2017 at (Formula presented) SR, while achieving consistently competitive SSIM performance across different datasets and scaling factors. These results demonstrate that IFNet effectively preserves fine anatomical structures and enhances reconstruction fidelity for medical image SR.<p></p> 2027-02-01T00:00:00Z Text Journal contribution 10.26187/deakin.34020477 https://figshare.com/articles/journal_contribution/Information_Filter-based_Diffusion_Transformer_for_medical_image_super-resolution/34020477 All Rights Reserved Restricted Access