Deep learning framework sharpens low-dose CT images while cutting noise

Low-dose computed tomography has become an increasingly common way to reduce the radiation a patient receives during scanning, but the dose savings come at a cost: images often carry excessive noise and degraded structure that can compromise diagnostic accuracy. A new study published in EJNMMI Physics on 16 September 2026 reports a deep learning framework, called ShearFocusNet, designed to suppress that noise while preserving the fine anatomical details that clinicians rely on.

The research, led by Qinglei Yan and Qiang Lin with colleagues including Xiaodi Huang, addresses two well-known weaknesses of existing deep learning denoising methods. The authors note that current approaches frequently over-smooth images and lack directional sensitivity, which makes it difficult for a network to keep thin edges, subtle textures and small anatomical structures intact once noise has been removed.

ShearFocusNet tackles the problem from two complementary directions: frequency and space. In the frequency domain, a Shearlet feature directional perception module uses adaptive multi-scale Shearlet transforms to capture edges, textures and noise structures with fine granularity. Shearlet transforms are well suited to this task because they can represent directional features at multiple scales, allowing the network to distinguish genuine anatomical contours from random noise patterns.

In the spatial domain, the framework introduces a dual-coordinate gated attention spatial focusing module. This module combines horizontal and vertical average pooling with max pooling through a gated attention mechanism, directing the network’s focus toward key features and strengthening its ability to discriminate between informative structure and unwanted noise. By pairing directional perception with coordinate attention-driven spatial focusing, the architecture aims to balance noise reduction against detail preservation.

The team evaluated the framework through extensive experiments on benchmark low-dose CT datasets. On the Mayo-2016 dataset, ShearFocusNet achieved a peak signal-to-noise ratio (PSNR) of 45.12 ± 0.42 dB and a structural similarity index (SSIM) of 97.57 ± 0.27%. On the Mayo-2020 dataset, the method reached 51.04 ± 0.18 dB PSNR and 99.31 ± 0.02% SSIM, while the Piglet dataset yielded 47.60 ± 0.19 dB and 99.17 ± 0.03%.

Beyond the headline PSNR and SSIM figures, the proposed method also recorded the highest edge preservation index (EPI) and the lowest gradient magnitude similarity deviation (GMSD) among the methods compared. Higher EPI values indicate better retention of edge information, while lower GMSD values reflect closer agreement of gradient structure with the reference, both of which matter when the goal is to keep diagnostically relevant detail visible after denoising.

According to the authors, these results confirm that ShearFocusNet can effectively suppress noise while preserving critical structural information. The study concludes that by uniting Shearlet transform-based directional perception with coordinate attention-driven spatial focusing, the framework achieves an effective balance between noise reduction and detail preservation, and the authors describe it as a robust and efficient approach that holds strong potential for advancing low-dose CT imaging in clinical practice.

As with any benchmark-driven deep learning study, the reported gains come from retrospective image-quality metrics on established datasets rather than from demonstrated diagnostic outcomes in patient care, so translating the framework into routine imaging workflows would require further validation. The article is published open access, and the version shared in September 2026 is a peer-reviewed, accepted manuscript carrying a permanent DOI, subject to further edits before the final version of record appears.

Subject of Research: A deep learning denoising framework for low-dose computed tomography images

Article Title: A deep learning framework with directional perception and spatial focusing for enhanced low-dose CT image denoising

Article References: Yan, Q., Lin, Q., Li, T., Fu, Y., Cao, Y., Liu, C., Cai, Z., & Huang, X. (2026). A deep learning framework with directional perception and spatial focusing for enhanced low-dose CT image denoising. EJNMMI Physics. https://doi.org/10.1186/s40658-026-00945-6

Image Credits: AI Generated

DOI: 10.1186/s40658-026-00945-6

Keywords: low-dose CT, deep learning, image denoising, Shearlet transform, attention mechanism, PSNR, SSIM, medical imaging, Mayo-2016 dataset, EJNMMI Physics, radiation dose reduction, deep