Clinical Assessment of CoLumbo Deep Learning System for Central Canal Stenosis Diagnostics

Abstract

Objectives: There is a great variability of inter-observer disagreements for central stenosis diagnostics depending on the used classification. This study investigates the level of agreement between lumbar magnetic resonance imaging (MRI) reports created by a deep learning neural network (CoLumbo) and the radiologists’ reading. Methods: A total of 382 (53.4 % females, 46.6 % males and average age 49.52±13.20) prospective consecutive patients in 3 different healthcare centers referred to L-spine MRI for back or leg pain were analyzed by the software CoLumbo for the presence of stenosis on all lumbar levels, by radiologists using it and radiologists not using the dedicated software. In case of disagreement between radiologists, a radiologist-arbiter opinion was used to establish majority opinion. The total number of evaluated levels was 1762. Results: There were 156 debatable cases of disagreements between radiologists using the software, and radiologists, not using CoLumbo, for the presence of central stenosis. In 18 cases, the arbiter opinion has coincided with that of the radiologist not using the software. In 138 cases, the former has coincided with that of the radiologist using the software CoLumbo. Most of the cases of disagreement are borderline cases. The reported sensitivity and specificity of CoLumbo was 92.70% and 99.04%, respectively. Conclusion: The study showed that the radiologist using the CoLumbo software achieved best results. The results of the algorithm were inferior but still better than radiologists not using the software in any published study.