Automated hip dysplasia detection using novel FlexiLBPHOG model with ultrasound images

Abstract

This study focuses on automatically detecting developmental hip dysplasia (DHD) using a novel feature engineering model, FlexiLBPHOG, inspired by the FlexiViT model. The model utilizes five patch types for feature extraction with local binary pattern (LBP) and histogram of oriented gradients (HOG) techniques. During feature extraction, five feature vectors are generated. In the next stage, three feature selection methods—Neighborhood Component Analysis (NCA), Chi-square (Chi2), and ReliefF (RF)—are used to select the top 500 features. Classification is performed using support vector machine (SVM) and k-nearest neighbors (kNN), resulting in 30 outcomes. Information fusion through iterative majority voting (IMV) and a greedy algorithm yields 58 outcomes, from which the best is selected. The FlexiLBPHOG model achieved a classification accuracy of 94.38% in detecting DHD in ultrasound images from a private dataset. The study confirms the effectiveness of the proposed model in image classification by integrating shallow image descriptors.

Keywords

• Development hip dysplasia detection
• FlexiLBPHOG
• Multiple feature selection
• Information fusion
• Ultrasound