Deep Convolutional Neural Networks Provide Motion Grading for High-Resolution Peripheral Quantitative Computed Tomography of the Scaphoid

In vivo high-resolution peripheral quantitative computed tomography (HR-pQCT) studies on bone characteristics are limited, partly due to the lack of standardized and objective techniques to describe motion artifacts responsible for lower-quality images. This study investigates the ability of such deep-learning techniques to assess image quality in HR-pQCT datasets of human scaphoids. In total, 1451 stacks of 482 scaphoid images from 53 patients, each with up to six follow-ups within one year, and each with one non-displaced fractured and one contralateral intact scaphoid, were independently graded by three observers using a visual grading scale for motion artifacts. A 3D-CNN was used to assess image quality. The accuracy of the 3D-CNN to assess the image quality compared to the mean results of three skilled operators was between 92% and 96%. The 3D-CNN classifier reached an ROC-AUC score of 0.94. The average assessment time for one scaphoid was 2.5 s. This study demonstrates that a deep-learning approach for rating radiological image quality provides objective assessments of motion grading for the scaphoid with a high accuracy and a short assessment time. In the future, such a 3D-CNN approach can be used as a resource-saving and cost-effective tool to classify the image quality of HR-pQCT datasets in a reliable, reproducible and objective way.