P29. Deep learning-based identification of lumbar spondylolysis on plain X-ray images using convolutional neural networks

Background Context

Lumbar spondylolysis is a common cause of low back pain in young patients, which is stress fracture of pars interarticularis due to excessive sports activity. Although most patients are treated by limitation of sporting activities and brace, accurate diagnosis is important because missed or delayed intervention can result in future spondylolisthesis. Based on plain X-ray imaging, it is difficult for the non-specialist to diagnose or decide to require an MRI scan as the next diagnostic step. Deep learning with the convolutional neural networks (CNNs) has attracted attention in the medical imaging field.

Purpose

To construct deep learning algorithms using CNNs to identify spondylolysis on plain X-ray images.

Study Design/Setting

Retrospective analysis using CT and plain X-ray images.

Patient Sample

This retrospective study included 100 patients with spondylolysis and 100 patients without spondylolysis under 20 years old.

Outcome Measures

We plotted the receiver operating characteristic curve and calculated the area under the curve (AUC) in order to evaluate the performance of the CNNs. Consequently, the sensitivity, specificity, and accuracy of the diagnosis by CNNs were calculated.

Methods

First, a digitally reconstructed radiograph (DRR) images were created from the CT image from 43 patients with spondylolysis, and the CNN was trained to identify vertebrae with and without spondylolysis on DRR images. Next, 100 vertebrae with spondylolysis and 100 normal vertebrae with matched levels were extracted from the anteroposterior and lateral plain X-ray images, respectively. The extracted images of 200 vertebrae were randomly divided into 150 vertebrae images for internal validation and 50 vertebrae images for external validation. Utilizing the trained model created from DRR images, fine tuning was conducted in training the plain X-ray images for internal validation. Finally, an external validation dataset was classified using 5 trained models created from plain X-ray images to validate the work of five models. In this study, a 25-layer CNN was used, including convolutional and pooling layers. Output information was binary classification regarding the presence or absence of spondylolysis. Five-fold cross-validation were conducted to assess the performance of CNNs. In addition, heatmaps of the CNNs focus site was created using the gradient-weighted class activation mapping method.

Results

In training using anteroposterior, X-ray images, five models of CNNs had performance with AUC of 0.82, sensitivity of 0.77, specificity of 0.80 and accuracy of 0.79 in internal validation, and AUC of 0.83, sensitivity of 0.77, specificity of 0.68 and accuracy of 0.74 in external validation. In training lateral plain X-ray images, five models of CNNs had higher performance with AUC of 0.97, sensitivity of 0.91, specificity of 0.96 and accuracy of 0.94 in internal validation, and AUC of 0.96, sensitivity of 0.87, specificity of 0.95 and accuracy of 0.90 in external validation. The high feature area on the heatmaps seemed to be around the pars interarticularis for both models.

Conclusions

This preliminary study showed high performance of deep learning algorithms with CNNs to identify lumbar spondylolysis on plain X-ray images. This model may help in the screening of lumbar spondylolysis in general practice without specialist for spine.

FDA Device/Drug Status

This abstract does not discuss or include any applicable devices or drugs.