Analysis of Developmental Dysplasia of the Hip Using Deep Learning Techniques

Abstract

Purpose: Developmental dysplasia of the hip (DDH) is a relatively common musculoskeletal condition in neonates. Early detection with ultrasound (US) is crucial for effective treatment. This study aimed to evaluate images obtained from hip ultrasonography with deep learning methods. Material and Method: Patients who underwent hip ultrasonography between January 2018 and September 2021 and were found to have normal hips and hip dysplasia were retrospectively screened. A total of 947 patient images, 450 girls and 497 boys, were examined. According to the Graf method, images were classified without any marking. In the first stage, two groups were created: those with Type 1 mature hips and those with dysplastic hips (other types). In the second stage of the study, four groups were created using only the α angle: 451 were classified as Type 1, 326 as Type 2a and 2b, 137 as Type 2c and D, and 33 as Type 3 and Type 4. During the classification, three versions of the EfficientNet model, one of the current deep learning models, were used. Classifiers were included in the study to improve the accuracy values of the models. In our study, two classifiers named support vector machine and K-nearest neighbors were used. Results: In the classification phase with deep learning models, the highest accuracy value of 0.9577 was obtained with the EfficientNetB1 model for 2 classes in the first group, while the highest accuracy value of 0.8571 was obtained with the EfficientNetB0 model for 4 classes in the second group. By including the classifiers in the evaluation, the highest accuracy rate was found to be 0.99 with EfficientNetB1 and 1(100%) with EfficientNetB2 in the first group, while it was 0.97 with EfficientNetB0 in the second group. Conclusion: In the diagnosis of developmental hip dysplasia, high accuracy rates were obtained in deep learning methods using US images. Accuracy rates increased with the addition of classifiers to the models. © 2025 Elsevier B.V., All rights reserved.