Classification of Spine X-ray Image using Neural Networks Models based on Haar Wavelet Features

The spine is a fundamental structure in the human body, playing a crucial role in supporting posture enabling individuals to stand upright, sit, walk, and engage in various activities. Maintaining spinal health is essential to ensure the body's optimal functioning and prevent disruptions to daily life. Among the disorders affecting the spine, scoliosis is characterized by abnormal lateral curvature, requiring early diagnosis to mitigate serious complications, such as chronic pain and respiratory dysfunction. This study examined the effectiveness of two Multilayer Perceptron (MLP) training algorithms—Gradient Descent with Adaptive Learning Rate (GDA) and Resilient Backpropagation (RP)—in classifying spinal X-ray images. Preprocessing steps included resizing, augmentation, and image enhancement through Average and Standard Deviation Stretching (AVSTDS) and Wiener filtering, while Haar wavelet transform was utilized for feature extraction. The models were evaluated using 10-fold cross-validation across varying hidden layer configurations (5, 10, and 15 layers). GDA consistently demonstrated superior performance, with the best results achieved in run 3 for five hidden layers (a training accuracy of 86.8% and a testing accuracy of 90.0%), run 7 for ten hidden layers (a training accuracy of 87.4% and a testing accuracy of 80.0%), and run 10 for 15 hidden layers (a training accuracy of 90.1% and a testing accuracy of 90.0%). Overall, GDA proved to be a more reliable and robust algorithm than RP for spine classification across all hidden layer configurations.