A hybrid deep learning approach based on u-net, alexnet and densenet for automatic detection of hip dysplasia

Objective: Developmental Hip Dysplasia (GCD) is an important orthopedic disorder that needs to be diagnosed early in the infantile period. In cases where early diagnosis is not made, serious motor function losses and complications requiring surgical intervention may occur. Current diagnostic processes are mostly based on clinical examination and ultrasonography. However, the user-dependent nature of these methods, the open nature of the interpretation differences and the lack of accessibility in every health center necessitates the development of artificial intelligence-based systems that can provide automatic and high accuracy diagnosis through X-ray images.The images are displayed in the literature, x-ray details It is also understood that most of the deep learning methods performed operate singularly and in this way production is limited. The main purpose of this programming is to develop a hybrid deep transfer learning model that can be detected using Developmental Hip Dysplasia with high accuracy rate of x-ray images. With this model, it is aimed to make a new contribution in the field of treatment image analysis and to obtain more reliable, faster and accurate results in DDH diagnosis. Materials and Methods: A three-stage hybrid deep learning architecture has been developed in the thesis study. In the first phase of the model, the U-Net architecture was used to segment low-level features from images. U-Net is an artificial neural network architecture that is often used in the field of medical image processing, especially successful in segmentation tasks. The feature map extracted at this stage was then transferred to the next stages for use for classification. In the second phase, DenseNet and AlexNet architectures are combined to learn more complex and abstract features. The intensive interlayer connectivity feature of DenseNet provides more efficient learning by reducing information loss. AlexNet is an architecture known for its classification success, especially in low-resolution images. By combining these two models, both a deep and compact feature extraction has been realized and the segmentation map from the first stage is fed into this structure. In the last stage, the output obtained from the combination of U-Net and DenseNet-AlexNet was combined and the classification process was realized. The model divides the images into two classes, "Displazi" and "Normal". For the training of the model, the open access data set used by Mohammad Fraiwan and colleagues in his work "Detection of developmental dysplasia of the hip in X-ray images using deep transfer learning" was used. This dataset consists of images of the pelvic anteroposterior X-ray of 4.5 infants (120 DDH, 234 normal) in the age group between 5 and 8 in total, shared with open access on the Kaggle platform, in order to evaluate the performance of the model during the test process In the data preprocessing phase, all images are standardized and pixels are normalized to optimize the learning process of the model. In the training process, the Adam optimization algorithm was preferred and a 5-fold cross-validation (K-Fold Cross Validation) method was used to increase the generalization of the model. Model performance was assessed by metrics such as accuracy (accuracy), precision (precision), specificity (specificity), and F1 score. Results: When the training and test results of the developed hybrid model were examined, it was seen that the model reached a high accuracy rate. At the end of the training process, the overall accuracy of the model was calculated as 97%. In addition, successful results have been achieved in other performance metrics such as sensitivity and specificity. Especially after the segmentation with the U-Net architecture, the combination of DenseNet and AlexNet architectures has increased the learning capacity of the model and positively affected the classification performance. The fact that the model gives consistent and successful results on different datasets shows that the generalability of the model is high and is suitable for real-world applicability. Conclusion: This study makes a significant contribution to existing literature by providing a hybrid deep learning approach for the detection of Developmental Hip Dysplasia from X-ray images. The three-stage hybrid model, achieved by combining U-Net, DenseNet and AlexNet architectures, offers high accuracy and reliability, and has the potential for a decision support system that can help physicians in the early diagnosis process. The findings show that this model is not only applicable in theoretical terms, but also in clinical settings. The use of larger data sets with future studies, the performance of the model can be further improved by the inclusion of different age groups and image sources. In addition, the model can be integrated into a real-time running software or mobile application, making it possible to use it directly in clinical applications. Thus, it can play an important role in the health sector as a tool that saves time and minimizes human error during the diagnostic process.