MRI of the spine is one of the commonest studies performed in clinical practice usually to study the cause of back pain. The reading of spine MR studies involves identifying the vertebral levels, estimating stenosis of the spinal canal, detecting the reduction in the vertebral height and also identifying spondylolisthesis. Several studies have shown that deep learning can assist in some of these tasks by automating them. In this study, we propose to use a platform-based approach for quick validation of a CNN based multi-modular set of custom Neural Networks that can automatically label the vertebral level and measure the central canal diameter, vertebral height, alignment, and disc height.
Methods and materials
The algorithm "Spindle" is a set of neural networks (developed by Synapsica technologies Pvt Ltd) that were trained on 11,321 Spinal MRIs acquired on different field strengths. The independent validation was done on 157 cases of lumbar spine MRI from 1.5 T and 3 T MRI scanners. The results of the algorithm were loaded on CARPL (CARING Analytics PLatform) and the spinal canal diameters at all five lumbar levels were plotted on Dot Plots with an adjustable threshold. Expert validation was done by a neuroradiologist with 9 years' experience. All levels where the network had an output of central canal diameter below 11mm were measured. The accuracy of vertebral labeling and listhesis detection was also verified.
Results
28 Slices with rotated images in the scans were excluded from the study. Out of remaining cases, there were a total of 49 slices fulfilling the threshold from 40 cases where the radiologist measured the spinal canal diameter. The mean percentage error of AI prediction of spinal canal diameter was 7.9%. There were a total of 16 spondylolisthetic levels, of which 12 were accurate. In two cases, there was correct identification but the defects were over graded. The vertebral labeling was accurate at all levels in these 40 cases.
Conclusion
The "Spindle" algorithm shows high accuracy in automatically labeling and quantifying several metrics of spondylosis in this retrospective study. Such automated solutions can enable faster reading and better quantification of imaging findings on MR scans.
ink to complete publication here: https://scholar.google.com/citations?view_op=view_citation&hl=en&user=LSnAkaUAAAAJ&cstart=20&pagesize=80&citation_for_view=LSnAkaUAAAAJ:Zph67rFs4hoC
