Detection of suspicious fractures and fissures with computer aided methods

X-ray images (radiographs) are among the most common ways to detect problems in bones and other organs of the human body. In some cases, depending on the experience of the radiologist, suspicious fractures can not be diagnosed correctly, and wrong treatments may be applied to the patient. For this reason, fracture / fissure detection, which is the first step of the treatment, is important. This study aims to enable the clinician to diagnose a fracture accurately and quickly, assist the clinician, and improve his/her performance in forensic cases or emergency rooms where access to experienced radiologists / orthopedists is limited. In addition, with this study, the performance of the EfficientNet architecture, which is among the state-of-the-art models developed in recent years, in classifying bone fractures has been investigated. Acccordingly, using the MURA data set, a deep learning application that can be accessed over the internet has been developed to detect fractures in the elbow, finger, forearm, hand, humerus, shoulder, wrist bones and to be used in the clinic. In the study carried out, models with ResNet50 and EfficientNet-B7 architecture were created with RF (random forest) and logistic regression (LR) classification algorithms. The accuracy rates of the created models in the MURA images and the images taken from the clinic were investigated. In the MURA data set, the accuracy rate was 82.4% in the elbow, 79.9% in the finger, 86.4% in the forearm, 83.5% in the hand, 86.5% in the humerus, 77.4% in the shoulder and 83.2% in the wrist images; a remarkable accuracy rate of 80.3% in the finger, 87.9% in the forearm, and 96.2% in the humerus images were obtained in clinical images.