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

Abstract

Low-dose computed tomography (CT) has been increasingly adopted to minimize radiation exposure. However, it often introduces excessive image noise and structural degradation, thereby compromising diagnostic accuracy. Existing deep learning-based denoising methods frequently suffer from over-smoothing and limited directional sensitivity, hindering their ability to preserve fine anatomical details. To overcome these challenges, we propose a novel framework called ShearFocusNet for denoising low-dose CT that leverages directional perception in the frequency domain and attention focusing in the spatial domain. Specifically, a Shearlet feature directional perception module employs adaptive multi-scale Shearlet transforms to capture edges, textures, and noise structures with fine granularity. We also introduce a dual-coordinate gated attention spatial focusing module that integrates horizontal and vertical average pooling and max pooling through a gated attention mechanism to enhance spatial focusing on key features, thereby enhancing the network’s discriminative capability. Extensive experiments on the benchmark low-dose CT datasets demonstrate that ShearFocusNet outperforms the advanced methods, achieving a PSNR of 45.12 ± 0.42 dB and an SSIM of 97.57 ± 0.27% on the Mayo-2016 dataset, 51.04 ± 0.18 dB and 99.31 ± 0.02% on the Mayo-2020 dataset, and 47.60 ± 0.19 dB and 99.17 ± 0.03% on the Piglet dataset. The proposed method also achieved the highest EPI and lowest GMSD. These results confirm that the proposed ShearFocusNet method can effectively suppress noise while preserving critical structural information. By uniting Shearlet transform-based directional perception with coordinate attention-driven spatial focusing, ShearFocusNet achieves an effective balance between noise reduction and detail preservation. This robust and efficient framework holds strong potential for advancing low-dose CT imaging in clinical practice.