Harnessing CNNs and Decision Trees for Autonomous Scoliosis Diagnosis on Spinal Radiographs

In recent decades, advances in fields such as robotics, computer graphics, and computer vision have led to an increase in interest in the study of image recognition among scientists. One of the challenges in classification of images is ensuring that the dataset used for training is sufficiently large and well-balanced to improve the effectiveness, robustness, and reliability of classification models. In this study, a research approach leveraging deep neural networks and decision tree models was employed to classify scoliosis severity levels, including mild, moderate, severe, and no scoliosis, with an emphasis on addressing class imbalance. By applying techniques to balance the dataset, significant improvements in classification accuracy were achieved. Specifically, the Xception convolutional neural network and XGBoost decision tree models achieved classification accuracies of $\mathbf{7 9. 2 5 \%}$ and $\mathbf{9 8. 6 1 \%}$, respectively. The approaches demonstrated high precision in categorizing scoliosis severity, offering a valuable tool for aiding clinical diagnosis and treatment planning.