Abstracts

RETURN TO ABSTRACT LISTING


2290. Hanging Protocol Optimization of Lumbar Spine X-rays with Machine Learning
Authors * Denotes Presenting Author
  1. Gene Kitamura *; University of Pittsburgh Medical Center (UPMC) and University of Pittsburgh
Objective:
The purpose of this study was to determine whether machine learning can be utilized to optimize the hanging protocol of lumbar spine radiographs. Specifically, we explored whether the trained models are able to accurately label lumbar spine views, detect hardware, and rotate the lateral views to straighten the image.

Materials and Methods:
We acquired approximately 3000 lumbar spine X-rays from a vendor neutral archive. We labeled each lumbar spine view as frontal (AP), right oblique (RO), left oblique (LO), right lateral (RL), left lateral (LL), right lumbosacral (RS), and left lumbosacral (LS). We also noted whether each X-ray had spinal hardware. For each of the lateral views, we noted the degree of rotation required to straighten the image. The data was split into a training, validation and test dataset. Utilizing augmentation, the image data was input into a Densenet-based model built with Tensorflow and Python. After training the model, the output metrics were evaluated on the test dataset. The objective function for the views and hardware presence was cross-entropy softmax. The objective function for the rotation value was mean squared error.

Results:
The area under the curve (AUC) for the receiver operator characteristic (ROC) curve was 0.99-1.00 for the lumbar spine views and hardware presence. For rotation of lateral views, the loss curve demonstrated expected decrease and plateau.

Conclusion:
Various research has shown the effectiveness of machine learning with musculoskeletal X-rays in detecting fractures (1–5), bone age (6), and arthritis (7). Specific to the lumbar spine X-rays, studies have shown that machine learning algorithms may have utility with predicting osteoporosis (8) and compression fractures (9). However, there is a relative paucity of research utilizing machine learning to improve the efficiency of current workflows. Routine radiology workflow could be streamlined with an optimized hanging protocol, such that all images are always presented in the same view order and with all lateral views corrected for rotation and facing to one side (left vs right). This method should be more effective than relying on the image metadata, which is prone to errors. Finally, being able to automatically detect hardware may be useful in automatic triaging for cross-sectional imaging protocols.