1119. Automated Multi-Modality Vertebral Deformity Assessment Using Deep Learning
Authors * Denotes Presenting Author
  1. Abhinav Suri *; University of Pennsylvania School of Medicine
  2. Grace Ng; University of Pennsylvania School of Medicine
  3. Brandon Jones; University of Pennsylvania School of Medicine
  4. Nancy Anabaraonye; University of Pennsylvania School of Medicine
  5. Patrick Beyrer; University of Pennsylvania School of Medicine
  6. Iman Fathali; University of Pennsylvania School of Medicine
  7. Chamith Rajapakse; University of Pennsylvania School of Medicine
Bone diseases cause an estimated 1.5 million individuals to suffer fractures each year in the United States alone. Of these fractures, vertebral fractures are among the most common, especially among individuals diagnosed with osteoporosis. However, only one-third of these fractures are clinically diagnosed. The process of doing so is time-consuming as three heights must be measured on each vertebral body. From these heights, vertebrae can be classified as either having a fracture (of type wedge, biconcave, or crush) or not (normal). Additionally, a grade (that quantifies severity) can be associated with each fracture using the Genant semiquantitative method. Deep learning algorithms can shorten the time to analyze imaging studies by detecting vertebral bodies in these studies and finding six landmarks on each vertebra that enables heights to be measured. The purpose of this study was to determine if such algorithms could both diagnose vertebral deformities and do so rapidly. We report the development of a neural network (type of deep learning algorithm) that can locate vertebral bodies and landmarks for making deformity diagnoses and report other clinically useful metrics (Lumbar Lordosis Angle; LLA) in less than 2 seconds per slice for MR, CT, and X-Ray imaging studies.

Materials and Methods:
Annotated sagittal spine MR (T1 & T2), CT, and X-Ray cases (MR=1123,age:67±11yrs; CT=137,age: 63±3yrs; X-Ray=484,age=64±2yrs) were used to train and test (80% and 20% of cases, respectively) three neural networks (a landmark detection network for each modality), using a variation of the Mask-RCNN architecture. Each case was manually annotated with six landmarks on each vertebral body (for height calculations). The networks were trained for 1000 epochs on the training dataset with the aim of minimizing the relative landmark error distance (i.e. X and Y distance between predicted and ground truth landmark locations relative to width and height of the vertebral body). Networks were evaluated using the testing dataset (which is not seen by the network during the training process) and evaluated by the aforementioned metric. L1-L5 LLA was calculated from landmarks predicted by the network (using the top corners of L1 vertebral body and bottom corners of L5 vertebral body) and correlated to the LLA calculated from manual annotations.

The network was able to predict relevant landmarks in <1.5 seconds. The network achieved a median relative landmark error distance of <4.6% across all modalities. It also achieved an overall accuracy of >91% for determining the deformity class and grade. The LLA predicted from the network also reached r>0.95 across all modalities when correlated to the calculated ground-truth LLA.

The neural network architecture was able to rapidly locate vertebral body landmarks for deformity diagnoses and measure LLA in MR, CT, and X-Ray imaging studies with high accuracy. This approach could simplify the screening and detection of changes in patients with vertebral deformities by reducing the burden on radiologists who have to do these measurements manually.