Automatic analysis of intervertebral discs based on deep learning – comparing preoperative MR images with 1-year post lumbar disc herniation surgery outcomes
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Degree project. Program in Medicine. Automatic analysis of intervertebral discs based on deep learning – comparing preoperative MR images with 1-year post lumbar disc herniation surgery outcomes . Emil Cedergårdh, 2019. Institute of Clinical Science, Department of Orthopedics. Gothenburg, Sweden. Introduction: Lumbar disc herniation surgery often leads to major improvement in leg pain; however, some patients have remaining back pain that might depend on the level of disc degeneration. Magnetic resonance imaging (MRI) examination has a central role in the preoperative evaluation. Today, the images are reviewed by a radiologist, a task which, in the future, might be assisted by computers using artificial intelligence (AI) and deep learning. In this study, intervertebral disc (IVD) characteristics from preoperative MR images is extracted and then compared with the 1-year post lumbar disc herniation surgery outcome. Due to technical error, semi-automated segmentation was used instead of deep learning-based segmentation. Aim: To study if there is a relationship between midsagittal signal intensity measures in preoperative MR images and 1-years postoperative patient reported outcome measures (PROM´s) on back pain, physical function and overall satisfaction. Method: Patients undergoing lumbar disc herniation surgery at Sahlgrenska University Hospital during the years 2013-2017 (n=218) and registered in the Swedish National Quality Registry for Spine Surgery (Swespine) were included. In each patient, the midsagittal part of the herniated IVD was segmented (outlined) on preoperative T2-weighted MR images using an in-house developed software. Signal intensity measures were calculated and statistically compared (t-test at p<0.05) to the PROM´s Numeric rating scale (NRS) back, Oswestry disability index (ODI) and Global Assessment (GA). Results: No significant difference in signal intensity measures between patients with successful versus unsuccessful PROMS´s was found. Conclusions: This study could not prove any relationship between midsagittal signal intensity measures in preoperative MR images and 1-years postoperative PROM´s. Further studies are encouraged using standardized MRI protocol and scanner, and more patient’s data enabling adjustment of confounders.