Diagnosis and classification of proximal femur fractures on hip radiographs with deep learning methods

4% of patients who have had a proximal femur fracture die in hospital, and 30% die within 1 year after treatment. The diagnosis of proximal femur fractures can be missed on hip radiographs due to the intensity of the emergency services and the short time allocated to the patients. Accurate and rapid diagnosis of proximal femur fractures and early initiation of treatment reduce patient morbidity and mortality. Four data sets were prepared. 178 x-ray images in our first dataset (VS-1), our second dataset (VS-2), VS-1 with CLAHE filter applied from image processing methods, in our third dataset (VS-3), VS-1 316 x-ray images and finally in our fourth data set (VS-4) were obtained by applying CLAHE filter to VS-3. Learning transfer models DenseNet201, Xception, ResNet152V2 and InceptionResNetV2 were used in the diagnosis of fractures. These models are trained on all datasets. F1 score of 0.9500 was achieved with the InceptionResNetV2 model in VS-1. F1 score of 0.9500 was achieved in DenseNet201 and InceptionResNetV2 models in VS-2. F1 score of 0.9000 was achieved in Xception and InceptionResNetV2 models in VS-3. In VS-4, F1 score of 0.9250 was achieved in the DenseNet201 model. In general, when the data set is increased and the CLAHE filter is applied from the image processing methods, the success rates of the models with low success increase. In this study, we have seen that when there is sufficient data and supported by image processing techniques, Artificial Intelligence technology can be used for optimized detection of proximal femur fractures and can also facilitate the work of clinicians.