Comparison of Lenke Classification using Multilayer Perceptron Based on AlexNet Feature Extraction Algorithm

The Lenke classification plays a crucial role in assessing the severity of scoliosis and guiding appropriate treatment decisions. Scoliosis is typically characterized by an abnormal lateral curvature of the spine and requires early detection to prevent complications such as chronic pain and respiratory issues. This study focuses on the classification of seven classes of spinal X-ray images, consisting of six Lenke types and one normal spine class. The study involves three main stages: pre-processing (including resizing, data augmentation, image enhancement using Average and Standard Deviation Stretching (AVSTDS), and Wiener filtering), feature extraction using the AlexNet architecture, and classification using a Multilayer Perceptron (MLP) model. Two MLP training model algorithms, Gradient Descent with Momentum (GDM) and Resilient Backpropagation (RP), were compared across three variations of hidden nodes (HN): 1, 5, and 10. The results indicate that RP outperformed GDM, particularly with 10 hidden nodes, achieving higher accuracy and more stable convergence.