Abstract
Abstract The ability to accurately diagnose and treat musculoskeletal diseases is a key function of medical imaging in healthcare. It is frequently difficult to interpret wrist X-rays because of slight anatomical differences as well as the requirement for quick decision making in clinical practice. While deep learning algorithms have demonstrated progress in medical image classification, previous research has mostly concentrated on a single architecture of convolutional neural networks (CNN) that may limit the variance of retrieved features and hence classification accuracy. To address this limitation, this study offers a hybrid deep learning system for automatic categorization of wrist anomalies using the MURA (XR_WRIST) dataset. The proposed methodology includes image resizing, grayscale conversion, Gaussian noise removal, normalization, contrast enhancement techniques using Adaptive Histogram Equalization (CLAHE) and class balancing techniques using Synthetic Minority Oversampling Technique (SMOTE). The deep features obtained from EfficientNetB7, ResNet101, MobileNetV2 and InceptionV3 were concatenated to take advantage of complementary representations of features and to enhance the classification results. The experimental results showed that the proposed hybrid model achieved a higher accuracy of 96.63%, precision of 98.96%, recall of 96.02%, F1 score of 97.46%, sensitivity of 96.02%, specificity of 97.91%, and AUC-ROC of 0.99 on the MURA wrist X-ray database than the individual backbone networks. Furthermore, ablation studies showed that the hybrid feature fusion strategy and SMOTE-based class balancing contributed to improved classification performance compared with individual models. The results show that the hybrid framework is a strong and efficient way to detect wrist abnormalities automatically, and it could help doctors make better decisions when analyzing musculoskeletal radiographs.