DeepSPINE: Automated Lumbar Vertebral Segmentation, Disc-level\n Designation, and Spinal Stenosis Grading Using Deep Learning

Abstract

The high prevalence of spinal stenosis results in a large volume of MRI\nimaging, yet interpretation can be time-consuming with high inter-reader\nvariability even among the most specialized radiologists. In this paper, we\ndevelop an efficient methodology to leverage the subject-matter-expertise\nstored in large-scale archival reporting and image data for a deep-learning\napproach to fully-automated lumbar spinal stenosis grading. Specifically, we\nintroduce three major contributions: (1) a natural-language-processing scheme\nto extract level-by-level ground-truth labels from free-text radiology reports\nfor the various types and grades of spinal stenosis (2) accurate vertebral\nsegmentation and disc-level localization using a U-Net architecture combined\nwith a spine-curve fitting method, and (3) a multi-input, multi-task, and\nmulti-class convolutional neural network to perform central canal and foraminal\nstenosis grading on both axial and sagittal imaging series inputs with the\nextracted report-derived labels applied to corresponding imaging level\nsegments. This study uses a large dataset of 22796 disc-levels extracted from\n4075 patients. We achieve state-of-the-art performance on lumbar spinal\nstenosis classification and expect the technique will increase both radiology\nworkflow efficiency and the perceived value of radiology reports for referring\nclinicians and patients.\n

Ähnliche Arbeiten