Deep Transfer Learning for Automated Multi-Class Classification of Spinal Disorders Using Benchmark and Clinical X-ray Datasets

Abstract

Among the most common spinal disorders are scoliosis, spondylitis, and spondylolisthesis. It is an essential reason for chronic pain and physical disability, and its accurate and timely diagnosis is essential. In contemporary years, deep learning technologies have demonstrated promising applications in automated medical image analysis. Despite this, the breadth and uniformity of publicly available data remain problems, as does analysis. Challenges for developing robust diagnostic models. This work proposes a deep transfer learning system that compiles spine X-ray data from benchmarks and hospitals to enhance the database, thereby standardizing it and improving classification performance. We used four state-of-art convolutional neural networks with a total of 131 million parameters. Four different network architectures, namely ResNet-152, DenseNet-201, Densenet-121, and VGG-16, have been trained, and the support was evaluated for both three and four-class spinal disorder classification. Storage and retrieval techniques. Data preprocessing and retrieval techniques. To mitigate overfitting and improve model inference, a method was employed. Accuracy, precision, recall, F1score, confusion matrices, and receiver operating characteristic (ROC) analysis were used to assess the model’s performance, with the area under the curve (AUC) serving as the main metric. The experimental results show the ResNet-152 consistently achieved higher accuracy than the other architectures, with a mean accuracy of 98.95% for three-class classification. In classification, the highest accuracy of 95.17% and 95.16% for four-class classification, with strong performance across precision, recall, F1-score, and ROC/AUC. The skip and residual learning strategy are responsible for the improved performance. Connection of ResNet-152, which enables efficient feature extraction and stable optimization in deep networks. Moreover, the inclusion of standardized benchmark and clinically obtained datasets improved model robustness and generalization across various spine diseases. The proposed framework provides a reliable computer-aided diagnostic approach that can help clinicians detect spinal diseases early and facilitate the development of functional artificial intelligence systems for medical imaging applications.