3D Metal Segmentation from few X-Ray Images for Metal Artifact Avoidance

Rohleder M, Kunze H, Maier A, Kreher B (2024)

Publication Language: English

Publication Type: Conference contribution, Abstract of a poster

Publication year: 2024

Journal

Progress in Biomedical Optics and Imaging - Proceedings of SPIE

Publisher: SPIE

Series: Proceedings of SPIE

City/Town: Bellingham, WA

Book Volume: 12925

Pages Range: 129253C

Conference Proceedings Title: Medical Imaging 2024: Physics of Medical Imaging

Event location: San Diego, CA US

ISBN: 9781510671546

DOI: 10.1117/12.3005260

Abstract

Purpose: Intraoperative Cone Beam CT (CBCT) is routinely used for implant placement verification but is confounded by metal artifacts obscuring the clinically relevant area around implants. This work reports a novel method to estimate a metal distribution from as few as two scout views which enables existing Metal Artifact Avoidance (MAA) methods to predict an artifact minimized scanning trajectory. Methods: A new model architecture was developed to directly predict a volumetric metal segmentation from given projection images. This is achieved by embedding a differentiable back projection operator into the model architecture which allows the network to transfer features extracted from projection images into the reconstruction domain. To ensure a generalization to clinical data, the model is trained on projection images from clinical and pre-clinical CBCT scans. The metal labels are generated by a previously trained model segmenting metal in the CBCT reconstructions. To counteract the limited available dataset sizes, a geometric data augmentation strategy is presented, which reflects transformations applied to the pixel data in the applied projection geometry. We evaluate the effects of this data augmentation and additionally assess the model's performance increase with respect to additional supplied views. Results: The newly introduced method achieves an average segmentation quality of 0.66±0.15 measured by the dice coefficient on the test set which is comparable to methods from literature requiring six views. The suggested data augmentation method has a strong regularizing impact on model convergence, resulting in improved generalization over the non-augmented trainings. When presented with a third view, the model's performance on average increases by 9.1% without re-training and 12.1% after finetuning. Conclusion: The presented method achieves competitive segmentation performance with only two scout images on realistic clinical data. This enables the application of existing MAA methods in the current clinical workflow without the need to acquire additional X-Ray scouts.

Authors with CRIS profile

Maximilian Rohleder Lehrstuhl für Informatik 5 (Mustererkennung) Andreas Maier Lehrstuhl für Informatik 5 (Mustererkennung)

How to cite

APA:

Rohleder, M., Kunze, H., Maier, A., & Kreher, B. (2024). 3D Metal Segmentation from few X-Ray Images for Metal Artifact Avoidance. Poster presentation at Medical Imaging 2024: Physics of Medical Imaging, San Diego, CA, US.

MLA:

Rohleder, Maximilian, et al. "3D Metal Segmentation from few X-Ray Images for Metal Artifact Avoidance." Presented at Medical Imaging 2024: Physics of Medical Imaging, San Diego, CA Ed. Rebecca Fahrig, John M. Sabol, Ke Li, Bellingham, WA: SPIE, 2024.

BibTeX: Download