Scholarlink

<p>Low-dose computed tomography (LDCT) imaging is widely used in clinical practice to reduce radiation exposure; however, LDCT images are inherently degraded by noise and streak artefacts, which can substantially impair diagnostic accuracy. Although deep learning-based denoising methods have shown promising results, many existing approaches generalize poorly across heterogeneous datasets and frequently introduce over-smoothing of critical anatomical structures, thereby limiting diagnostic reliability, robustness, and clinical trust. To address these challenges, this study proposes ED-SCMA, a novel Encoder-Decoder architecture with Skip Connections and a Multi-Scale Attention Module for accurate and robust LDCT image denoising. The proposed framework integrates dual parallel spatial attention modules (PSAM), a multi-scale convolutional network (MSCN), and residual learning within a symmetric encoder-decoder design. PSAM adaptively refines fine-grained anatomical details while capturing global contextual dependencies, whereas MSCN enhances feature representation across multiple spatial scales, preserving both local textures and global structural information. Residual skip connections further stabilize feature propagation and recovery, explicitly mitigating the over-smoothing of vital anatomical organs. The proposed method is rigorously evaluated on multiple benchmark and clinical datasets, including the AAPM pelvic CT dataset, ACR phantom data, and hospital-acquired maxillofacial, paranasal sinus, and temporal bone datasets. Across all datasets, ED-SCMA consistently outperforms state-of-the-art denoising methods on standard image quality metrics. Specifically, ED-SCMA achieves PSNR/SSIM/RMSE values of 41.08/0.947/0.0040 on pelvic CT data, 44.19/0.948/0.0028 on ACR phantom data, 30.53/0.680/0.0141 on maxillofacial data, and 27.99/0.624/0.0175 on sinus and temporal bone datasets. Comprehensive ablation studies further demonstrate that an encoder-decoder depth configuration of 7:3 combined with dual PSAM provides the optimal balance between denoising performance and computational efficiency. Overall, ED-SCMA represents a methodologically original and clinically significant contribution to LDCT image reconstruction. By jointly addressing generalization, anatomical fidelity, and computational efficiency, the proposed approach advances the state of the art in medical image denoising and supports safer, higher-quality CT imaging for routine clinical deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ED-SCMA: encoder-decoder with skip-connection and multi-scale attention module for low-dose CT denoising

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Temitope Emmanuel Komolafe,
  • Liang Zhou,
  • Adegbola Oyedotun Adeniji,
  • Blessing Funmi Komolafe,
  • Jingzhen Guo,
  • Asad Saleem,
  • Wenlong Zhao,
  • Oluwarotimi Williams Samuel

摘要

Low-dose computed tomography (LDCT) imaging is widely used in clinical practice to reduce radiation exposure; however, LDCT images are inherently degraded by noise and streak artefacts, which can substantially impair diagnostic accuracy. Although deep learning-based denoising methods have shown promising results, many existing approaches generalize poorly across heterogeneous datasets and frequently introduce over-smoothing of critical anatomical structures, thereby limiting diagnostic reliability, robustness, and clinical trust. To address these challenges, this study proposes ED-SCMA, a novel Encoder-Decoder architecture with Skip Connections and a Multi-Scale Attention Module for accurate and robust LDCT image denoising. The proposed framework integrates dual parallel spatial attention modules (PSAM), a multi-scale convolutional network (MSCN), and residual learning within a symmetric encoder-decoder design. PSAM adaptively refines fine-grained anatomical details while capturing global contextual dependencies, whereas MSCN enhances feature representation across multiple spatial scales, preserving both local textures and global structural information. Residual skip connections further stabilize feature propagation and recovery, explicitly mitigating the over-smoothing of vital anatomical organs. The proposed method is rigorously evaluated on multiple benchmark and clinical datasets, including the AAPM pelvic CT dataset, ACR phantom data, and hospital-acquired maxillofacial, paranasal sinus, and temporal bone datasets. Across all datasets, ED-SCMA consistently outperforms state-of-the-art denoising methods on standard image quality metrics. Specifically, ED-SCMA achieves PSNR/SSIM/RMSE values of 41.08/0.947/0.0040 on pelvic CT data, 44.19/0.948/0.0028 on ACR phantom data, 30.53/0.680/0.0141 on maxillofacial data, and 27.99/0.624/0.0175 on sinus and temporal bone datasets. Comprehensive ablation studies further demonstrate that an encoder-decoder depth configuration of 7:3 combined with dual PSAM provides the optimal balance between denoising performance and computational efficiency. Overall, ED-SCMA represents a methodologically original and clinically significant contribution to LDCT image reconstruction. By jointly addressing generalization, anatomical fidelity, and computational efficiency, the proposed approach advances the state of the art in medical image denoising and supports safer, higher-quality CT imaging for routine clinical deployment.