Deep learning for the radiographic diagnosis of proximal femur fractures: Limitations and programming issues

Radiology is one of the domains where artificial intelligence (AI) yields encouraging results, with diagnostic accuracy that approaches experienced radiologists and physicians. Diagnostic errors in traumatology are rare but can have serious functional consequences. Using AI as a radiological aid may be beneficial emergency room. Thus, an effective, low-cost software helps making radiographic diagnoses would relevant tool for current clinical practice, although this concept has rarely been evaluated orthopedics proximal femur fractures (PFF). This led us to conduct prospective study goals of: 1) programming deep learning help make diagnosis PFF on radiographs 2) evaluate its performance. It possible program effective based limited number radiographs. Our database consisted 1309 radiographs: 963 had PFF, while 346 did not. The sample size was increased 8-fold (resulting 10,472 radiographs) using validated technique. Each radiograph by orthopedic surgeon RectLabel™ (https://rectlabel.com), differentiating between healthy fractured zones. Fractures were classified according AO system. algorithm programmed Tensorflow™ (Google Brain, Santa Clara, Ca, USA, tensorflow.org). In all, 9425 annotated (90%) used training phase 1074 (10%) test phase. sensitivity 61% femoral neck 67% trochanteric fractures. specificity 69%, positive predictive value 55% 56%, negative 74% 78%, respectively. results not good enough our practice. Programming sufficient only done several tens thousands radiographs, or transfer learning. III; studies, Study nonconsecutive patients, without consistently applied reference “gold” standard.