Using Deep Learning to Enhance Reporting Efficiency and Accuracy in Degenerative Cervical Spine MRI.

Abstract

Background context: Cervical spine MRI is essential for evaluating degenerative cervical spondylosis (DCS) but is time-consuming to report and subject to interobserver variability. The integration of artificial intelligence in medical imaging offers potential solutions to enhance productivity and diagnostic consistency.

Purpose: To assess whether a transformer-based deep learning model (DLM) can improve the efficiency and accuracy of radiologists in reporting DCS MRIs.

Study design/setting: Retrospective study using external DCS MRIs from December 2015 to August 2018.

Patient sample: The test dataset comprised 50 pre-operative DCS MRIs (2,555 images) from 50 patients (mean age = 60 years ± SD 14; 13 women (26%)), excluding cases with instrumentation.

Outcome measures: Primary outcomes were interpretation time and interobserver agreement (Gwet's kappa) among radiologists grading spinal canal and neural foramina stenosis with and without DLM-assistance.

Methods: A transformer-based DLM was used to classify spinal canal (grades 0/1/2/3) and neural foramina (grades 0/1/2) stenosis at each disc level. Two experienced musculoskeletal radiologists (both with 12-years-of-experience) provided reference standard labels in consensus. Ten radiologists (0-7 years of experience) graded DCS MRIs with and without DLM-assistance, with a 1-month washout period between sessions to minimize recall bias. Interpretation time and interobserver agreement were assessed.

Results: DLM-assistance significantly improved interpretation time by 69-308 seconds (p < 0.001), reducing mean time from 159-490 seconds (SD 27-649) to 90-182 seconds (SD 42-218). Radiology residents experienced the largest time savings. DLM-assistance improved interobserver agreement across all stenosis gradings compared to baseline. For dichotomous spinal canal grading, residents had the largest improvement in agreement (κ = 0.63 to 0.77, p < 0.001). Conversely, for dichotomous neural foramina grading, musculoskeletal radiologists had the largest improvement (κ = 0.60 to 0.72, p < 0.001). Notably, independent DLM performance alone was equivalent or superior to all readers.

Conclusions: The integration of a deep learning model into the radiological assessment of DCS MRI improved radiologists' interpretation time and interobserver agreement, regardless of experience level.