Hounsfield Unit Preservation in Low‐Dose Abdomen and Pelvic CT Denoising: An Enhanced UNIT Framework for Radiotherapy Planning

ABSTRACT

The clinical utility of deep learning for low‐dose CT (LDCT) denoising is often limited by a failure to preserve the Hounsfield Unit (HU) values that are essential for radiotherapy planning. Many models prioritize visual quality over the quantitative accuracy required for dose‐sensitive workflows. This study presents an enhanced framework designed to address this critical gap. We propose a paired image‐to‐image translation framework based on the UNIT architecture, incorporating a dedicated HU preservation loss. The model was implemented in MATLAB's Deep Learning Toolbox and trained on paired abdominal and pelvic LDCT/HDCT scans. Performance was evaluated using standard image quality metrics (PSNR, SSIM), voxel‐wise HU Mean Absolute Error (MAE), and an extended ROI‐based analysis across clinically relevant tissue classes spanning the full HU range, including air cavities, adipose tissue, liver, contrast‐enhanced vessels, and dense bone. The proposed model significantly outperformed baseline methods, achieving a mean PSNR of 43.83 dB, SSIM of 0.968, and MAE of 11.5 HU on abdominal scans. The ROI analysis confirmed high quantitative fidelity across tissue types, with mean HU deviations remaining below 3 HU over the full evaluated HU spectrum and near‐zero error observed in both soft tissue and dense bone regions. By integrating an explicit HU preservation constraint, the proposed framework denoises LDCT images while maintaining quantitative accuracy across diverse tissue densities. The method establishes a benchmark for HU‐consistent LDCT enhancement and supports clinical translation by producing quantitatively consistent, DICOM‐compliant CT outputs suitable for downstream validation in radiotherapy treatment planning systems.