Improving Radiographic Fracture Recognition Performance and Efficiency Using Artificial Intelligence

Abstract

MUSCULOSKELETAL IMAGINGF racture detection using radiography is one of the most common tasks in patients with high-or low-energy trauma in various clinical settings, including the emergency department, urgent care, and outpatient clinics such as orthopedics, rheumatology, and family medicine. Missed fractures on radiographs are one of the most common causes of diagnostic discrepancies between initial interpretations by nonradiologists or radiology residents and the final read by board-certified radiologists, leading to preventable harm or delay in care to the patient (1-3). Fracture interpretation errors can represent up to 24% of harmful diagnostic errors seen in the emergency department (2). Furthermore, inconsistencies in radiographic diagnosis of fractures are more common during the evening and overnight hours (5 pm to 3 am), likely related to nonexpert reading and fatigue (3). In patients with multiple traumas, the proportion of missed injuries, including fractures, can be high on the forearm and hands (6.6%) and feet (6.5%) (4,5).To date, several studies about artificial intelligence (AI) aid to fracture detection have been performed focusing only on certain body parts, such as hand, wrist, and forearm (6-9); hip and pelvis (10,11); knees (9); and spine (12). One study evaluated fractures in 11 body locations, Background: Missed fractures are a common cause of diagnostic discrepancy between initial radiographic interpretation and the final read by board-certified radiologists.Purpose: To assess the effect of assistance by artificial intelligence (AI) on diagnostic performances of physicians for fractures on radiographs. Materials and Methods:This retrospective diagnostic study used the multi-reader, multi-case methodology based on an external multicenter data set of 480 examinations with at least 60 examinations per body region (foot and ankle, knee and leg, hip and pelvis, hand and wrist, elbow and arm, shoulder and clavicle, rib cage, and thoracolumbar spine) between July 2020 and January 2021. Fracture prevalence was set at 50%. The ground truth was determined by two musculoskeletal radiologists, with discrepancies solved by a third. Twenty-four readers (radiologists, orthopedists, emergency physicians, physician assistants, rheumatologists, family physicians) were presented the whole validation data set (n = 480), with and without AI assistance, with a 1-month minimum washout period. The primary analysis had to demonstrate superiority of sensitivity per patient and the noninferiority of specificity per patient at 23% margin with AI aid. Stand-alone AI performance was also assessed using receiver operating characteristic curves.Results: A total of 480 patients were included (mean age, 59 years 6 16 [standard deviation]; 327 women). The sensitivity per patient was 10.4% higher (95% CI: 6.9, 13.9; P , .001 for superiority) with AI aid (4331 of 5760 readings, 75.2%) than without AI (3732 of 5760 readings, 64.8%). The specificity per patient with AI aid (5504 of 5760 readings, 95.6%) was noninferior to that ...

Automated fracture screening using an object detection algorithm on whole-body trauma computed tomography

“…With the assistance of a CNN model, orthopedic surgeons can achieve an improved sensitivity for fracture detection and a reduced the time to read and interpret CT scans. In a previous fracture classification study of 480 patients, CNN model assistance for radiographic reading by six types of readers (emergency physicians, orthopedists, radiologists, physician assistants, rheumatologists, family physicians) showed a 10.4% improvement in fracture (thoracolumbar spine, rib cage, hip and pelvis, shoulder and clavicle, elbow and arm, hand and wrist, knee and leg, and foot and ankle) detection sensitivity (0.752 vs 0.648; p < 0.001 for superiority) without a reduction in specificity (0.956 vs 0.906, p = 0.001 for noninferiority), and the diagnosis time was shortened by an average of 6.3 s per patient with CNN model assistance ( p = 0.046) 21 . Furthermore, Sato et al developed a CNN classification model from a relatively large dataset of hip fractures on plain radiographs.…”

Can artificial intelligence pass the Fellowship of the Royal College of Radiologists examination? Multi-reader diagnostic accuracy study

“…14 In another study, artificial intelligence was also shown to improve radiologists' sensitivity by 10.4% (versus without artificial intelligence) for interpreting musculoskeletal radiographs and to reduce reporting time by 6.3 seconds per examination. 15 Nevertheless, the artificial intelligence should analyse only radiographs that it has been trained to interpret to avoid erroneous anomalies being highlighted on those deemed non-interpretable, which occurred in one case in this study (and has been known to occur with other artificial intelligence tools). 16 Interestingly, in this study, radiologists slightly overestimated the likely performance of the artificial intelligence candidate, assuming that it would perform almost as well as themselves on average and outperform them in at least three of the 10 mock examinations.…”

Neural Architecture Search for Fracture Classification

“…All the authors de facto reuse an off-the-shelf architecture designed and pretrained on datasets of natural images, which they fine-tune on private annotated datasets of fractures. This work is based on previous studies of Gleamer [4,6], in which a Mask-RCNN [7] model pre-trained on COCO, is assembled using the De-tectron2 framework [23] and fine-tuned on a private internal dataset of 60,000 radiographs of patients with trauma gathered from 22 institutions and annotated by medical experts.…”

“…A promising direction to facilitate interpretation and reduce the prevalence of errors is to assist radiologists with a computer-aided diagnosis software. The recent advent of deep learning has made such software capable of improving the performance of radiologists [9] and even outperform experts on their own [4,6].…”