1579. Fully Autonomous Delineation of Neural Foramina and Central Canal Anatomy in Lumbar Spine MRIs Using Machine Learning

Authors: Mark A. Attiah, MD; Bilwaj Gaonkar, PhD; Diane Villaroman, BS; Yasmine Alkhalid, BS; Steve Moran, BS; Christine Ahn; Matthew Edwards, BS; Joel Beckett; Noriko Salamon, MD; Alex Bui, PhD; Luke Macyszyn (Los Angeles, CA)

Introduction: Narrowing of the neural foramina and spinal canal in the lumbar spine leads to neurogenic claudication and lumbar radiculopathy. There are currently no quantitative tools to objectively measure these anatomical regions. Hence, we aimed to develop automated analytical techniques to quantify imaging markers in order to perform a disciplined analysis of the relationship between imaging markers and symptomatology. Methods: We developed a two-stage, fully automated learning based algorithm to segment and measure both neural foramina and spinal canals using T2-weighted MR images. First, an ensemble of support vector machine classifiers are trained to differentiate windows that contain the structure of interest from those that do not. Then, an ensemble of regression trees precisely delineate the anatomy of interest within this window. Training data is generated from manual labeling provided by expert physicians. We compare the segmentations generated by the machine algorithm to those generated by two human raters in 100 sagittal and axial MR images. Results: We found that the proposed ensemble techniques achieve close to human accuracy in segmenting spinal canals using axial MR images, as measured by a Dice score of 0.84. Similarly, the proposed methodology achieves a Dice score of 0.63 in segmenting neural foramina on sagittal images. We also demonstrate that even a state of the art deep network achieves a Dice score of 0.61. Conclusion: We developed a machine learning technique to detect and segment foramina and spinal canals on MRI in less than a minute. These imaging markers may be used for diagnosis, disease monitoring and aid clinical decision making in symptomatic spinal stenosis and radiculopathy.