Differentiating Types of Spinal Curvatures Using AlexNet and ResNet-50 Models

Scoliosis is a deformity of the spine characterized by a lateral curvature, often in an $S$ or C shape. Scoliosis is usually diagnosed through a physical examination and using a combination of X-rays, CT scans, or MRIs. However, it is possible to misdiagnose scoliosis, especially if the curvature is mild or if the doctor is inexperienced. Therefore, a breakthrough in the early detection of scoliosis is highly required. AI has excellent potential in detecting scoliosis from X-rays. There have been many studies that utilized AI in diagnosing spinal disorders, and most of them obtained excellent results. Hence, it is necessary to compare models and methods to determine the best from them. This study applied two pre-trained models from ImageNet: AlexNet and ResNet-50. This study discovered a slight difference in the average accuracy obtained by both models. The difference was only 1.25% in training and 0.12% in testing. In training and testing measurements, ResNet obtained a perfect score of 100%. However, in terms of time measurement, it required four times more than AlexNet. ResNet-50 took 8 minutes and 58 seconds for training, while AlexNet only took 2 minutes and 4 seconds. These measurements and tests concluded that the two models had excellent performance. Hence, they could help researchers and those who need it in spine image classification.