In the context of traumatic emergencies, a considerable number of musculoskeletal radiographs are performed every day. They are the first line imaging modality to diagnose bone fracture, due to their availability, speed, cost, and low radiation [1,2]. The interpretation of trauma radiographs is a demanding task that requires radiologic expertise and time, unfortunately sometimes resulting in missed fractures, a preventable cause of morbidity [3]. Indeed, fracture is the most common condition in the emergency department that results in delayed or erroneous diagnosis [4]. Extremity fractures are the second most frequently missed diagnosis leading to a claim, after breast cancer [5]. Litigation in medicine is becoming more frequent in developed countries [6], and the financial burden of litigations in trauma is often in paradox to the severity of the injury itself [7].

The initial applications of artificial intelligence (AI) in medicine focused largely on image-recognition diagnostic tasks such as detecting mammographic lesions and recognizing skin cancer [8,9,10,11]. Data challenges, organized by the French Society of Radiology, came up with innovative AI solutions for the current relevant problems in radiology, such as classification of benign or malignant breast nodules on ultrasound examinations, detection and contouring of pathological neck lymph nodes from cervical computed tomography (CT) examinations and classification of calcium score on coronary calcifications from thoracic CT examinations [10]. The developing use of AI tools in radiology has led to discussions conducted by a national working group, in order to suggest rules of good practice for using these tools [12]. The World Health Organization also published principles to ensure AI works for the public interest in all countries [13].

Recent studies showed promising results of the use of AI for detecting bone fractures [14], [15], [16], [17], [18], especially peripheral skeletal fractures, which allowed algorithms to obtain Food and Drug Administration approval and CE marking as a class IIa medical device [19,20]. A recent study also showed the reliability of a deep learning algorithm for detecting fractures in children [21]. A recent meta-analysis that compared the diagnostic performance in fracture detection between AI and clinicians pinpointed that included studies only focused on the performance of AI regarding diagnosis of fracture but did not evaluate the impact of AI on the radiologist's workflow [22].

The main objective of this study was to compare the diagnostic performance between unassisted radiologists, AI-assisted radiologists and stand-alone AI, for the detection of bone fractures. Secondary objective was to evaluate the impact of the use of an AI algorithm in the radiologist's daily trauma emergencies workflow.