Bone fracture detection and classification using node-level capsule graph neural network with X-ray images of broken and unbroken bones
D Kanagaraj, Antonio Oliver, R. Ranjith, J. N.
Abstract
Bone fractures are frequent and need to be diagnosed rapidly to be treated appropriately. Detecting fractures is challenging due to the variability in bone size and shape. The small size of certain bones and their intricate structures make detection more difficult. In this manuscript, Bone Fracture Detection and Classification Using Node-Level Capsule Graph Neural Network with X-ray Images (BDC-NCGNN-XRI-BUB)are proposed. Bone fracture images are initially collected from a standard dataset and pre-processed using the Distributed Nonlinear Polynomial Graph Filter (DNPGF) to reduce noise and enhance contrast. These enhanced images are then processed using the High-Order Synchroextracting Transform (HOST) to extract key textural features such as contrast, correlation, homogeneity, energy, and dissimilarity. The extracted features are input into a Node-Level Capsule Graph Neural Network (NCGNN) for classifying images as fractured or non-fractured. To improve classification accuracy, NCGNN is optimized using the Gazelle Optimization Algorithm (GOA). The proposed BDC-NCGNN-XRI-BUB model is implemented in MATLAB, and performance is evaluated using metrics like Accuracy, Precision, Recall, F1-score, Error rate, and ROC. The BDC-NCGNN-XRI-BUB approach’s performance is 18.75%, 26.89%, and 32.57% more accurate; 16.87%, 24.57%, and 32.94% more precise; and 18.43%, 25.64%, and 31.40% more recall when observed using existing methods like Bone Fracture Classification in X-Ray Using Deep Learning Models (VDN-CMIS-BFD), and Detection of bone fracture based on machine learning techniques (DBF-MLT) and Diagnosis Model for Bone Fracture Identification of Athlete from X-Ray and MRI Images (HSDS-DM-BFI-XRMI), methods respectively.