Musculoskeletal disorders, which affect the locomotor system including the spine, have become a major medical, social and economic problem that, because of its high prevalence and the increasing number of patients, unfavorably impacts the quality of life. Radiological examination of spine images has a crucial role in surgery planning and establishing treatment strategies for many musculoskeletal and spinal disorders. Orthopedics, the medical specialty that focuses on the musculoskeletal system, has embraced imaging as an integral part of diagnosis, treatment and follow-up. By interpreting image information, different quantitative features can be extracted that help in image-assisted orthopedic examinations, such as the evaluation and prediction of the geometrical characteristics of the spinopelvic complex. Recently, advances in artificial intelligence (AI) have been recognized as valuable in tasks related to imaging analytics, and with the increase in computational power and general availability of data in the past decade, a major leap in the performance has been observed with the advent of deep learning (DL). In the proposed research project, we will design, develop and evaluate intelligent AI-based imaging analytics algorithms that aim to improve medical image interpretation, and investigate their integration with clinical practice in the field of orthopedic imaging and management of musculoskeletal and spinal disorders. In orthopedics, imaging analytics is mostly concentrated around the segmentation of spine structures and measurement of spinopelvic parameters. Spine segmentation represents the localization and delineation of the boundaries of individual vertebrae in the given image, while spinopelvic parameter measurement is focused on image-assisted evaluation of scoliosis, kyphosis, lordosis and sagittal balance as clinical expressions of the spinal curvature and body posture. For the purpose of the proposed AI-based imaging analytics, we will first devise a database consisting of radiographic (X-ray), computed tomography (CT) and magnetic resonance (MR) spine images with corresponding reference annotations in the form of vertebral segmentation masks and spinopelvic parameter measurements. By building on our previous work, we will then develop state-of-the-art DL algorithms for spine segmentation and modeling from CT and MR images, and landmark detection and modeling from X-ray images. As measurements are considered to be more reliable when extracted from three-dimensional (3D) CT or MR images, they will be transferred to two-dimensional (2D) X-ray images by 3D-2D mapping. Afterwards, the spinopelvic parameters will be measured from X-ray images by relying on the accurate information obtained by 3D spine segmentation of CT and MR images, resulting in a complete radiological analysis of the spinopelvic complex. To augment the obtained results, we will then focus on the evaluation of the interpretability of the developed DL algorithms, obtained through uncertainty estimation that describes the confidence of an AI method in its output. Finally, we will devise a software framework to merge the developed DL algorithms with tools for medical image manipulation, and ensure their end-to-end functionality. The methodological advances and the obtained results will be disseminated by publication in top-ranking scientific journals and presentation at international conferences, as well as by means of open science and scientific networking. Our final goal is therefore to devise a comprehensive software framework consisting of AI-based imaging analytics with integrated interpretability evaluation that will enhance medical image interpretation, facilitate the clinical workflow from the perspective of radiological examinations of the spine, improve the quality of life of patients with musculoskeletal disorders, and enable further research on innovative orthopedic applications.
