Background/Purpose: Diffuse idiopathic skeletal hyperostosis (DISH) is a non-inflammatory condition most classically seen in the spine, and is characterized by ossification of the spinal ligaments and entheses. DISH may be asymptomatic, but in some cases is associated with back pain and stiffness. Risk factors include male sex, increasing age, obesity and metabolic syndrome. Important radiographic mimics of DISH include osteoarthritis and ankylosing spondylitis There are multiple classification criteria for DISH, though the 1976 Resnick and Niwayama classification criteria are best known (1).

While xrays are used to identify DISH, it can still be challenging for trainees and primary care health professionals to distinguish it from its mimics. We conducted a pilot study to develop an Artificial Intelligence (AI) assistant, trained to identify DISH from ‘non-DISH’ on xrays by using convolution neural network (CNN). CNN is the most popular model for imaging classification in deep learning.

Methods: DISH patients were identified from our university’s radiology database, using word search ‘DISH’ or ‘Diffuse idiopathic skeletal hyperostosis’ in all cervical, thoracic or lumbosacral radiographs in the year 2015. Diagnosis of DISH was confirmed by 2 board certified radiologists, and either completely fulfilled Resnick criteria or incompletely fulfilled Resnick criteria but were still felt to have DISH based on radiologists’ expert opinion. We also compiled a list of age and sex-matched patients with spinal x-rays from 2015 who lacked DISH based on the radiologists’ read.

All image data were imported into R statistical language/Rstudio. We used a CNN to train a classification rule for spinal xrays of DISH vs. non-DISH using the keras and tensorflow R packages. 90% of each group was used as training data and the rest were used as test data to evaluate the accuracy of the classification rule.

Results: 116 patients with DISH and 262 matched controls were included. The image size was cropped from 601×524 pixels to 461 x 345 pixels to remove unnecessary solid background and add-on information before the CNN training. The trained CNN was a sequential model with 2 convolution layers (64 or 32 rectified linear units), 2 dense layers (one layer with 16 rectified linear units and the other with 1 sigmoid unit), and others. The network was trained by the RMSprop optimizer and validation split = 0.2 with 100 epochs. The overall training accuracy of the classification rule by CNN was 92% (sensitivity 75% and specificity 100%). The overall test accuracy was 89% (sensitivity 70% and specificity 98%).

Conclusion: The trained AI-assistant has reasonable sensitivity and very high specificity. In this pilot study, we demonstrated that a CNN-based AI-assist for identifying DISH xrays could be trained from scratch with a moderate size of images that seemed to perform reasonably well. The future research will be focusing on improving the accuracy of the AI-assistant with more images.

« Back to ACR Convergence 2020

ACR Meeting Abstracts - https://acrabstracts.org/abstract/an-artificial-intelligence-ai-assistant-identifying-spinal-diffuse-idiopathic-skeletal-hyperostosis-on-plain-x-rays-a-pilot-deep-learning-study/