BoneView

Evaluation of the medical impact of artificial intelligence for limb and pelvic bone fracture detection

Regnard et. al

RSNA

Published In

RSNA (2021)

Authors

Nor-Eddine Regnard, Boubekeur Lanseur; Louis Lassalle; Antoine Feydy; Nicolas Theumann.

Abstract

Purpose :

 

To compare the performance of routine radiological interpretation and artificial intelligence (AI) in the detection of limb and pelvic bone fractures and assess its impact in terms of changes in medical care. 

 

Materials and methods :

 

We retrospectively evaluated all X-ray exams performed for a suspected limb or pelvic bone fracture during 3 consecutive months (Jan.-Mar. 2017) in a private imaging group of 17 centers. Each exam was analyzed by the AI BoneViewTM and its results were compared to those of the initial report. In case of discrepancy, the exam was reviewed by a senior skeletal radiologist (12 y. of experience) to settle on the presence of one or several fractures. He specified whether a complementary imaging exam was necessary following the change in diagnosis and evaluated the possible consequences of a delay in therapeutic care (0: none, 1: low, 2: impact on limb function, 3: vital).

 

Results : 

 

A total of 4774 exams (0-103 y.o., 30% pediatrics, 49% women) were included in the study. The anatomical locations of the exams were: hand (22%), wrist (4%), foot (13%), ankle (11%), leg (5%), knee (23%), arm (3%), elbow (2%), shoulder (9%), pelvis (8%). There were a total of 703 discordant exams reviewed by the senior radiologist. Of the 785 positive exams, 903 fractures were detected by either the AI and the initial report together (n=651, 72.1%) or by the senior radiologist in discordant cases (n=252, 27.9%). Of these 252 fractures, the senior radiologist confirmed the presence of 235 fractures detected by the AI and not mentioned in the initial report (26%) and of 17 fractures (1.9%) present in the initial report but not pointed by the AI. Among these 235 fractures, 46 were second fractures. Of the 235 additional fractures detected by the AI, 69 (29.4%) should have led to another imaging exam (CT or MRI) to confirm the diagnosis (n=61, 26%) or to characterize the severity of the fracture (n=48, 20.4%). The knowledge of the fracture could have led to a change in therapeutic care for 144 fractures (61.3%) and a possible surgery for 51 (21.7%). The consequences of a delay in therapeutic care were evaluated as low (score 1) for 155 fractures (66%) and as a source of impact on limb function (score 2) for 24 fractures (10.2%). Of the 235 additional fractures, 32 were impactful fractures: 6 pelvises, 5 scaphoids, 4 tibial plateau, 16 distal radii, and 1 fibula neck. 

 

Conclusion :

 

On a clinical routine dataset with natural prevalence, the AI detected 235 additional fractures with 46 multi-fracture exams, corresponding to an increase of 26% in fracture detection. Overall, 5% of exams were subject to medical errors. The best performances were obtained by the radiologist + AI. 

 

Clinical Relevance Statement/Application : 

 

AI-aid has the potential to prevent medical errors like missed fractures on initial X-ray exams thus impacting patient care.