AI MSK clinical applications: spine imaging

Abstract

Recent investigations have focused on the clinical application of artificial intelligence (AI) for tasks specifically addressing the musculoskeletal imaging routine. Several AI applications have been dedicated to optimizing the radiology value chain in spine imaging, independent from modality or specific application. This review aims to summarize the status quo and future perspective regarding utilization of AI for spine imaging. First, the basics of AI concepts are clarified. Second, the different tasks and use cases for AI applications in spine imaging are discussed and illustrated by examples. Finally, the authors of this review present their personal perception of AI in daily imaging and discuss future chances and challenges that come along with AI-based solutions.

Keywords: Artificial intelligence; Spine.

Fig. 1

The radiology value chain, as described by Enzmann [2]. The main components of a classical complete radiology service model are image acquisition, “read” images, report, and medical decision

Fig. 2

Routinely implemented automated image processing in spine imaging. Complex reconstructions of the whole spine are already implemented in clinical workflows for degenerative disease as well as in the trauma setting. The left part of the image shows automated output for spine labelling which allows for angle-corrected axial analysis of the vertebral structures (e.g., joint facets). Correct angulation for L5/S1 joint was detected (circled in red). The right part of the image summarizes standard workup of the spine and ribs with stretched multiplanar reconstructions as routinely done in trauma patients at the authors’ institute. Whereas ribs can be easily noted as normal, the compression fracture of the fourth lumbar vertebra (white arrowheads) is not only easy to detect, but can also be rapidly assessed with regard to complicating factors, e.g., spinal stenosis or instability. Images were acquired in Siemens syngo.via (syngo.via VB30A Bone reading, Siemens Healthineers, Erlangen, Germany)

Fig. 3

Example images from deep learning image segmentation in whole-body MRI. The images represent coronal multiplanar reconstructions of a T1-weighted Dixon-based dataset of a healthy individual. From left to right, fat and gadolinium-enhanced water sequences, as well as manually segmented “ground truth” segmentation mask and its automatic “pendant,” predicted by a deep learning–based MRI segmentation algorithm. Red, green, and blue areas represent the compartments subcutaneous adipose tissue, visceral adipose tissue, and muscle mass, respectively. All images unpublished own data, copyrighted by the authors

Fig. 4

Algorithm performance expressed as probabilities of nerve position as either root, trunci, or fascicles in sagittal MR images of the brachial plexus (own unpublished data). Image numbers increase from medial to lateral, beginning at the cervical spine (3-mm slice thickness)

Fig. 5

Quantitative (top row) vs. qualitative (bottom row) assessment of lumbar spinal stenosis severity. Texture analysis (TA) proved excellent reproducibility and objectivity regardless of whether only the central spinal canal was assessed (top left, red), or if instead the epidural sac and lateral recesses (top red, yellow) were included for measurements of the cross-sectional area. Moreover, TA outperforms qualitative approaches that differentiate between severe (bottom left) and extreme cases with epidural fat obliteration (bottom right)