Advances in Daily Musculoskeletal Imaging: Automated Analysis of Classic Radiographs

See also the article by Schock et al in this issue.

Gustav Andreisek, MD, MBA, is a general radiologist with a special dedication to magnetic resonance and musculoskeletal imaging. His interests include innovations in the field of medical imaging as well as their transition into clinical routine. Dr Andreisek aims for sustainable leadership and structural developments in radiology. His main position is chief of radiology at a public hospital in Switzerland.

Long-leg radiography is a well-established, decades-old radiographic examination that is a mainstay in daily patient care (1). The examination is frequently performed before total hip, knee, or ankle arthroplasty in degenerative disease (2,3). Long-leg radiographs also may be requested postoperatively as a marker for surgical outcome (4,5). Other less frequent indications include congenital, growth-related, and posttraumatic disorders. Many orthopedic surgeons aim for a radiographic evaluation of the whole lower extremity at once as long-leg radiographs allow various measurements such as of the (vertical) axis of the legs. Typically, long-leg radiographs are taken in upright, weight-bearing, and standing positions and include both legs (6–8). Most radiography systems acquire two to three slightly overlapping radiographs ranging from the pelvis to the foot. Proprietary software stitches them together to one full overview radiograph which is then sent as a standard Digital Imaging and Communications in Medicine image into the picture archiving and communication system.

In this issue, Schock et al present an artificial intelligence (AI)–based method for automated analysis of long-leg radiographs (9). The authors used 255 long-leg bilateral radiographs which were obtained during a 9-month period at their institution. They separated the imaging data into training, validation, and test datasets to teach their AI algorithms to measure the (vertical) leg axis on long-leg radiographs in a similar way as a human being would do it in daily clinical routine. Standard of reference was manual and software-assisted measurements performed by two radiology residents. The AI algorithm was based on a convolutional neural network and allowed measurement of the so-called hip-knee-ankle angle and the femoral anatomic-mechanical angle. Results showed that both measurements could be achieved accurately and reliably across a wide range of morphologic-anatomic configurations, irrespective of the presence of orthopedic hardware. In addition, Schock et al noted that the measurements were performed more rapidly than manual measurements (9). The AI algorithm was indeed 10- to 40-fold faster.

This well-executed study showed an inherent limitation of AI algorithms: namely, that they may fail in certain circumstances. Schock et al stated that their “automatic segmentation was not perfect and [was] challenged by orthopedic hardware, extremity shape and position, joint degeneration, bone texture and structure, and image quality” (9). Therefore, radiologists should not trust AI algorithms blindly, and results should always be double-checked. In selected cases, manual measurements may still be necessary. However, overall, this study has shown the already high accuracy and reliability of AI-based image evaluation. This study well mimicked the daily clinical routine of radiologists. Therefore, I fully agree with the author's conclusion that their solution might be ready for use in a real clinical scenario and that it has the potential to provide radiologists relief from a time-consuming routine task (9). Of course, further research is needed to improve some weaknesses of the algorithm as addressed in this article. In addition, prior to any clinical use, clearing by the responsible authorities as a medical device is usually required. Especially the latter will be an issue for many AI developments in the future as they emerge now in all fields of radiology, and since they continue to prove their readiness for clinical use. The current publication by Schock et al is an important step in the right direction, and such articles are the building blocks required to establish AI in routine clinical practice.