Author: Edward Robert Criscuolo, Chenlu Qin, Deshan Yang, Zhendong Zhang 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University 🌍
Purpose:
Low-dose CT (LDCT) imaging minimizes radiation exposure but introduces significant noise, compromising image quality. While deep learning-based denoising models such as HFormer achieve state-of-the-art performance, their implementations are often complex and designed for 2D slice-by-slice processing using patch operations, which can disrupt intra-slice consistency. This work extends HFormer for direct 3D volumetric denoising, eliminating the need for slice-wise processing. Additionally, we introduce an easy-to-use MATLAB and command-line tool, making deep learning-based LDCT denoising more accessible to researchers and clinicians.
Methods:
We developed a streamlined MATLAB framework that facilitates efficient 3D LDCT denoising by directly processing volumetric medical images. The HFormer pipeline was modified to accept 3D image volumes as input and output fully denoised 3D volumes, preserving continuity across slices. Our MATLAB framework automates pre-processing, HFormer execution, and post-processing, ensuring seamless integration into existing medical imaging workflows. To address block artifacts in 3D volumes, we optimized the patch-wise processing strategy, improving intra-slice consistency and reducing reconstruction artifacts. The proposed 3D denoising approach was evaluated across multiple datasets covering different anatomical sites and image qualities to assess its generalizability.
Results:
Our method enables direct 3D LDCT denoising, significantly reducing intra-slice artifacts and improving spatial consistency compared to conventional 2D-based approaches. The optimized volumetric processing strategy results in smoother reconstructions with preserved structural integrity. Additionally, our tool features a user-friendly interface, making it suitable for clinical deployment, with potential for real-time integration into medical image viewers.
Conclusion:
We present an accessible, efficient, and enhanced 3D LDCT denoising framework based on HFormer, effectively overcoming the limitations of 2D slice-wise processing. Our MATLAB-based tool bridges the gap between research and clinical implementation, providing a practical and scalable solution for LDCT noise reduction.