Abstracts

RETURN TO ABSTRACT LISTING


E2374. Deep Learning in Spine Health Assessment
Authors
  1. Abhinav Suri; University of Pennsylvania
  2. Anjali Gupta; University of Pennsylvania
  3. Vikram Balasubramanian; University of Pennsylvania
  4. Winnie Xu; University of Pennsylvania
  5. Anita Kalluri; University of Pennsylvania
  6. Charis Ma; University of Pennsylvania
  7. Chamith Rajapakse; University of Pennsylvania
Background
Bone diseases are highly prevalent across the world. In the United States alone, nearly 1.5 million individuals suffer fractures due to bone diseases each year. Among individuals diagnosed with osteoporosis, nearly 50% of them suffer a vertebral fracture at one point in their life, leading to decreased mobility and increased mortality. Thus, it is necessary to find ways to monitor spine health to intervene before major fracture events occur. One way to do so is by assessing imaging studies for vertebral deformities and overall bone health; however, the process of making measurements necessary to diagnose individuals with mild, moderate, and severe deformities is time-consuming, contributing to the fact that only one-third of fractures are clinically diagnosed and vertebral deformities are under-reported at rates as high as 85% among radiologists. Deep learning algorithms offer solutions that can be applied to the world of medical imaging. Some of these algorithms can detect objects, find landmarks in each object, and produce segmentations delineating the exact pixel locations and the border of each object. This exhibit will focus on explaining these algorithms and showing how they can be adapted for detecting vertebral body deformities and measuring other bone quality metrics.

Educational Goals / Teaching Points
At the end of this presentation, the learner should have knowledge of the following: 1. Semiquantitative method for vertebral deformity assessment (wedge, biconcave, and crush) 2. How neural networks can be used for feature detection in MR, CT, and X-Ray scans 3. How Landmark detection networks can be used for automated vertebral height quantification and deformity diagnoses 4. How segmentation networks can be used to extract vertebral bodies and volumes 5. How these automated measurements can be used for vertebral bone quality assessment

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
This exhibit will focus on the diagnosis of vertebral deformities per the Genant semi/fully quantitative method, specifically how these diagnoses can be automated using artificial neural network architectures that detect six relevant landmarks on each vertebral body. From these landmarks, three heights (anterior, Ha; middle, Hm; posterior, Hp) can be measured and deformities can be classified according to type and severity based on reductions in each height. Furthermore, we will show the potential application that segmentation of vertebral bodies and discs can have in diagnostics and research. The accuracy of these neural network designs will be shown in MR, CT, and X-Ray imaging studies.

Conclusion
We will show how the neural network architectures can rapidly locate vertebral body landmarks for height calculations and produce segmentation masks across MR, CT, and X-Ray images with high accuracy. Furthermore, we will explain how this approach could simplify the screening and detection of changes in patients with vertebral deformities and fractures, reducing the burden on radiologists who have to do these measurements manually.