Motion grading of high-resolution quantitative computed tomography supported by deep convolutional neural networks.

Lay Summary

When patients move during a specific type of bone scan called high-resolution peripheral quantitative CT (HR-pQCT), it reduces the quality of the images. This makes it harder to precisely measure bone density and structure. Normally, operators have to manually inspect the images and score any motion artifacts - a tedious process. Images with too much motion may be incorrectly accepted for analysis if missed on initial review. This leaves patients with potential inaccurate scan results. Novel artificial intelligence methods called a convolutional neural network (CNN) can classify images and detect patterns. Therefore, this study aimed to develop a CNN that automatically scores motion from HR-pQCT scans. The CNN also indicates when manual review is still needed due to uncertainty. Our CNN scored motion in seconds with high accuracy comparable to operators. It had good precision in detecting true motion, good recall in detecting all motion present, and strong agreement with operator scores. This AI post-processing tool could significantly reduce operator time spent assessing HR-pQCT image quality. It can be quickly implemented to evaluate how motion impacts measured bone properties. Ultimately, it has potential to improve consistency and quality control for clinical HR-pQCT scanning.

Recommended citation: Walle M, Eggemann D, Atkins PR, Kendall JJ, Stock K, Müller R, Collins CJ. Motion grading of high-resolution quantitative computed tomography supported by deep convolutional neural networks. Bone. 2023 Jan 1;166:116607. https://www.sciencedirect.com/science/article/pii/S8756328222002848