From synthesis to clinic: a task-specific framework for CT images metal artifact reduction in cervical brachytherapy

Abstract

Severe metal artifacts in cervical brachytherapy CT scans obscure critical anatomical structures and applicator geometry. Existing metal artifact reduction (MAR) models are inadequate in addressing these complex artifacts, leading to uncertainties in organs at risk (OARs) delineation and dose optimization that may compromise treatment safety and precision. This study aims to develop and validate a task-specific, synthetic-to-clinical deep learning framework for MAR in cervical brachytherapy CT scans. The synthetic dataset consisted of 500 external beam radiotherapy CT scans with simulated applicator artifacts from March 2021-April 2025. An independent clinical validation cohort included 50 brachytherapy CT scans. A proposed task-specific framework was applied to baseline deep learning MAR models, including InDuDoNet+, OSCNet+, OSCNet, and ADN. The performances of these models were compared against standard linear interpolation (LI) method and uncorrected images. Performance was assessed using quantitative metrics (peak signal-to-noise ratio [PSNR] and structural similarity index measure [SSIM]) on synthetic data, and clinical metrics (artifact index, applicator reconstruction error, and geometric accuracy of OARs) on clinical data. Additionally, three experienced radiologists independently evaluated the image quality using a 5-point Likert scale. For the synthetic dataset, the proposed task-specific models significantly outperformed baseline models, with improvements of 32.42% in SSIM and 35.67% in PSNR ( P < .001). Among the evaluated models, the task-specific InDuDoNet+ demonstrated superior performance (SSIM, 0.97; PSNR, 36.66 dB). In the clinical validation of 50 patients, task-specific models achieved significantly higher expert ratings (4.5 vs. 3.5 for uncorrected images) and exhibited an approximately 50% reduction in artifact indices for the bladder and rectum. Besides, the task-specific InDuDoNet+ model reduced the applicator reconstruction error to 0.1 mm, improved geometric contouring accuracy (Hausdorff distances < 1.0 mm for both bladder and rectum), and shortened contouring time by 30% to 40%. This study developed a task-specific deep learning framework to reduce brachytherapy CT metal artifacts while preserving anatomical details. The superior performance of the task-specific InDuDoNet+ model in improving geometric accuracy and workflow efficiency indicates that integrating this technology into clinical practice could significantly enhance the precision of cervical cancer brachytherapy. This trial was registered at the Chinese Clinical Trial Registry (ChiCTR) (registration number: MR-14-26-010321; date of registration: 2025-06-12).