Prediction of Spinal Abnormalities in Neuroradiology Images Applying Deep Transfer Learning

The vertebral column is the structure that provides support, protection, and facilitates human mobility. There are vertebral column deformities such as scoliosis and spondylolisthesis that can cause discomfort in patients. Therefore, early detection of these deformities through radiological images is of great importance to prevent their progression. Nowadays, there are systems that use artificial intelligence to efficiently classify these images. In this case, deep transfer learning with pre-trained models was employed to classify radiographic images into three different classes (no scoliosis, scoliosis, and spondylolisthesis). The research followed a quantitative approach, had an exploratory scope, and an experimental design. The methodology implemented was iterative through four increments. The first increment focused on testing and fine-tuning the MobileNetV3-Small, MobileNetV3-Large, and EfficientNetV2B0 models. In the second increment, tests and fine-tuning were conducted on the EfficientNetV2B1, EfficientNetV2B2, and InceptionResNetV2 models. In the third increment, a database of medical images from Honduran patients was created. In the fourth, predictions were made using the best-performing models. As a result, it was found that in the first increment, an average accuracy of 97.01% was achieved, and in the second increment, an average accuracy of 98.01% was reached for the evaluated models. The creation of a local database for future research lines was consolidated, and the classification algorithm was validated through predictions of radiographic images of Honduras patients.