Deep Learning Dramatically Reduces the Work Associated with Image Cataloguing and Analysis

Abstract

Commentary Artificial intelligence (AI), machine learning, neural networks, and shallow and deep learning are all forms of data analytics that are creeping into the orthopaedic lexicon1—with good reason, as these forms of data analytics have also become prevalent across medicine and in many aspects of life outside of medicine and health care. It is imperative that the field of orthopaedics take advantage of tools that have demonstrated benefit in other related applications in order to advance orthopaedic research and ultimately point-of-care decisions. The article by Rouzrokh et al. describes and evaluates a process employing a form of AI known as deep learning to identify and catalogue hip images from an uncatalogued institutional image warehouse. In essence, the study seeks to determine whether a computer can do what has previously been the tedious manual work of clinicians and investigators seeking to identify and measure images for research purposes. There are a number of aspects to this manuscript that make it an excellent read. First of all, the authors provide readers with tangible evidence that deep learning can be applied with confidence to identify images of certain descriptions for registry or quality-improvement purposes. Next, reading this manuscript will help orthopaedic surgeons to understand the language and meaning of deep learning. Additionally, this paper demonstrates how a deep-learning process can be utilized to both identify and perform simple measurements on a large collection of images from a data warehouse. Consider that validation of the AI process, which was performed as part of the methods of this analysis, necessitated human review of 5,000 images. The actual automated AI analysis took less than 8 hours to perform. Although the authors do not state how much time the human review of images required, one can safely assume that it was far greater than the 8 hours required by the algorithm to identify and measure 846,988 DICOM (digital imaging and communications in medicine) files. Thus, the article demonstrates the massive reduction in work that this technology offers. The authors offer the possibility that the 3-step pipeline to create their tool could be utilized "by other institutions or registries to construct radiography databases for patient care, longitudinal surveillance, and large-scale research. The stepwise approach for establishing a radiography registry can further be utilized as a workflow guide for other anatomic areas." These comments and other reports in the literature clearly suggest that the technology applied to this total hip arthroplasty radiographic registry is possibly applicable to the work of investigators in other subspecialties2. Likewise, there is a community of investigators utilizing deep-learning tools who are prepared to share their experience and expand the use of these tools more broadly in our profession. Finally, as one looks ahead at the future and what will grow from the introduction of this technology into musculoskeletal science, it is clear that we are only at the beginning of what is possible. The authors have done an excellent job of demonstrating how this technology may be utilized to perform the simple tasks of image identification and angular measurement with an acceptable degree of accuracy. One can only imagine that more sophisticated image identification and measurement are possible by developing similar systems with greater maturity, sophistication, and experience. Deep learning could be utilized to provide other types of data that would be useful both in clinical investigation and at the point of care, such as markers of implant loosening, wear, or osteolysis, or measurement of skeletal morphology from more sophisticated and advanced imaging3. There are endless possibilities for what might be made possible by such robust analytics in our new digital and connected world.

Ähnliche Arbeiten