Determining hip dysplasia using deep learning methods

Developmental hip dysplasia (DDH) is a disease in which the hip joint fails to develop normally due to various reasons before, during or after birth. In studies conducted in our country, it has been observed that dislocation rates are as high as 5-15:1000. The most important method used to detect developmental hip dysplasia (DDH) is hip ultrasonography (US). The stage of obtaining the hip US image varies because it depends on the operator and external influences. This variability may lead to misclassification. In this thesis, an artificial intelligence based system has been developed to eliminate this variability and minimise errors. The developed system includes a 2-stage deep learning model. In the first stage, U-NET architecture is used and in the second stage, region-based convolutional neural network architecture with mask is used. Firstly, it is checked whether the hip US image is taken correctly, that is, whether it can be analysed. It is checked whether the 3 basic anatomical structures (ischium, ilium, labrum) that need to be determined are detected. Then, the parallelism of the iliac bone to the baseline is checked. If the ilium is parallel to the baseline and all anatomical structures are detected, the US image is classified as analysable. Alpha and beta angles are calculated on this US image classified as analysable and developmental hip dysplasia is classified. It is thought that the system developed in this thesis will increase the accuracy of the diagnosis to be made by eliminating inter-operator variability in the diagnosis of developmental hip dysplasia and thus will help the experts