Author: Rani Anne', Wenchao Cao, Yingxuan Chen, Wookjin Choi, Firas Mourtada, Yevgeniy Vinogradskiy 👨🔬
Affiliation: Thomas Jefferson University 🌍
Purpose: In-room mobile cone-beam CT (CBCT) is emerging to enhance high-dose-rate (HDR) brachytherapy workflow using on-demand imaging. However, metal artifacts from X-ray markers inside gynecological (GYN) applicators significantly degrade image quality. We developed MARCI (Metal Artifact Reduction for CBCT Imaging), a self-supervised deep learning framework, to reduce these artifacts and improve organs-at-risk (OARs) delineation—specifically the bladder, rectal wall, and vaginal mucosa—critical for HDR planning.
Methods: MARCI leverages the spatial localization of metal artifacts in brachytherapy. The framework consists of three stages: (1) Artifact Classification: A convolutional neural network classifies slices as containing "Severe Artifacts" or "Minimal Artifacts" (trained on 100 manually labeled slices). (2) Unsupervised Artifact Reduction: A CycleGAN trained on 591 severe and 605 minimal artifact slices enhances image quality through unsupervised learning. (3) Supervised Refinement: A U-Net, trained with 500 high-quality, manually selected CycleGAN outputs paired with their original counterparts, performs supervised artifact correction, mitigating potential hallucination artifacts. The system was developed using 1,196 image slices from seven GYN HDR patients imaged on an Elekta ImagingRing CBCT. Quantitative evaluation of HU accuracy was performed on 280 uniform regions of interest (ROIs) from 10 slices reserved for testing.
Results: The artifact classification network achieved an AUC of 0.92 (95% CI: 0.85–0.98) on a balanced test set (n = 100). Quantitative analysis of 280 ROIs revealed improved HU consistency, with MARCI images achieving a mean HU of 57.50 compared to 74.41 in original images. Reference images without markers showed an HU of 60.80. Qualitative assessments confirmed MARCI significantly reduced artifacts while preserving anatomical details.
Conclusion: MARCI effectively reduces metal artifacts in CBCT images using a self-supervised approach that does not require paired training data. Enhanced imaging holds significant potential for improving OAR contouring accuracy, especially in tissues adjacent to GYN applicators, leading to more precise HDR brachytherapy treatment planning.