Research on AI classification of thoracolumbar fractures based on deep convolutional neural network and transfer learning

Abstract

The traditional thoracolumbar fracture image-assisted classification method has low accuracy and poor generalization ability. Therefore, based on deep convolutional neural network, this paper proposes an AI classification method for thoracolumbar fracture for auxiliary diagnosis. Firstly, a total of 698 CT images of patients with lumbar spine fractures were collected from Sichuan Integrative Medicine Hospital, and a data set was established, including 279 compression fractures (category A), 295 burst fractures (category B), and 124 normal (category C). Secondly, the convolutional neural network model ResNet-50 was modified and combined with transfer learning to train, verify and test the data set to obtain the AI classification model of thoracolumbar fracture. Then, the Confusion Matrix is used to evaluate the prediction model. The accuracy of the training set and the validation set of the model is 95.75% and 96.36%, respectively, indicating that the model obtained by training has good accuracy and generalization ability. This paper proposes an image-assisted classification method for thoracolumbar fracture, which can improve the efficiency and accuracy of manual diagnosis.

Keywords

• thoracolumbar fracture;deep convolutional neural network;AI classification method;generalization ability