2254. Rapid Quantification of Spine Metrics in Multiple Imaging Modalities Using Deep Learning
Authors * Denotes Presenting Author
  1. Abhinav Suri *; University of Pennsylvania
  2. Sungho Kim; University of Pennsylvania
  3. Yi-An Tu; University of Pennsylvania
  4. Thomas Leichner; University of Pennsylvania
  5. Grace Choi; University of Pennsylvania
  6. Nikita Bastin; University of Pennsylvania
  7. Chamith Rajapakse; University of Pennsylvania
Vertebral fractures comprise nearly 50% of all fractures in patients diagnosed with osteoporosis. These fractures are quantified by measuring three heights (anterior; Ha, middle; Hm, and posterior; Hp) on each vertebral body in an imaging study. However, the process of doing so is time-consuming, leading to under-reporting rates of vertebral deformities as high as 85% amongst radiologists. Additionally, imaging studies can yield useful information for researching bone quality via the analysis of vertebral body characteristics if given a segmentation mask that delineates the precise location of vertebral bodies (which is also time-consuming to manually segment). Deep learning algorithms can shorten the time to analyze imaging studies by automatically extracting imaging features. The purpose of this study was to determine if such algorithms both could output spine metrics (vertebral body heights and segmentation masks) that are useful for assessing spine health and do so rapidly. We report the development of a neural network (type of deep learning algorithm) that can locate vertebral body landmarks for height calculations and perform segmentations for imaging studies 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±3 yrs; X-Ray=484, age: 64±2 yrs) were used to train and test (80% and 20% of cases, respectively) six neural networks (a landmark and segmentation network for each modality), using a variation of the Mask-RCNN architecture. Each case was manually annotated with landmarks on each vertebral body (for height calculations) and a segmentation mask. The networks were trained for 1000 epochs (landmark networks) or 600 epochs (segmentation networks) 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 vertebral body) and maximizing the 2D Dice score (ranges from 0 to 1 where 1 means predicted segmentation = manually labeled segmentation in a single slice). Networks were evaluated using the testing dataset (which is not seen by the network during the training process) using the aforementioned metrics.

The network was able to predict relevant landmarks and segmentation masks in < 1.6 seconds. The network achieved a median relative landmark error distance of < 4.6% across all modalities in the X and Y direction. Predicted Ha, Hm, and Hp had a < 1% error relative to the original height. The network also achieved an overall 2D Dice score > 0.95 (all modalities).

The neural network architecture was able to rapidly locate vertebral body landmarks for height calculations and produce segmentation masks across MR, CT, and X-Ray images with high accuracy. This approach could simplify the screening and detection of changes in patients with vertebral deformities and fractures by reducing the burden on radiologists who have to do these measurements manually.