Related Experiment Video
Updated: May 30, 2025
05:17
Author Spotlight: Scope of LE-ULBD as a Safe, Effective, and Minimally Invasive Approach to Treat Lumbar Spinal Stenosis
Published on: February 9, 2024
505
Artificial intelligence for segmentation and classification in lumbar spinal stenosis: an overview of current methods
E J A Verheijen1,2, T Kapogiannis3, D Munteh3
1Computational Neuroscience Outcomes Center, Department of Neurosurgery, Brigham and Women's Hospital, Harvard Medical School, Boston, USA. e.j.a.verheijen@lumc.nl.
Summary
Machine learning models show excellent performance in segmenting and classifying lumbar spinal stenosis (LSS). Deep learning methods, particularly U-Net, outperform conventional approaches, but standardization and external validation are needed for broader clinical use.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Spinal Diagnostics
Background:
- Lumbar spinal stenosis (LSS) is a common condition caused by degenerative narrowing of the spinal canal.
- Current diagnostic methods rely on manual interpretation of imaging scans (MRI, X-ray, CT), which is time-consuming and prone to variability.
- A lack of standardized grading systems contributes to inter-reader variability in LSS assessment.
Purpose of the Study:
- To systematically review existing machine learning (ML) models for the segmentation and classification of lumbar spinal stenosis (LSS).
- To identify the current landscape of ML applications in diagnosing LSS severity.
- To understand the performance and limitations of various ML algorithms in LSS analysis.
Main Methods:
- A systematic literature review was conducted across major databases (Cochrane Library, Embase, Emcare, PubMed, Web of Science).
- Studies describing ML algorithms for lumbar spine segmentation or classification for LSS were included.
- Risk of bias was assessed using an adapted Newcastle-Ottawa Scale, and qualitative analysis focused on algorithm type (conventional ML vs. Deep Learning) and task.
Main Results:
- 27 articles were included, focusing on segmentation (9), classification (16), or both (2), primarily using MRI data.
- Machine learning algorithms demonstrated excellent performance for both segmentation and classification tasks.
- Deep learning (DL) models, especially U-Net architectures for segmentation and U-Net/CNNs for classification, generally outperformed conventional ML methods.
Conclusions:
- Deep learning models show superior performance over conventional ML for LSS segmentation and classification.
- Direct comparisons between studies are difficult due to diverse outcome measures and datasets.
- Future research should prioritize DL for segmentation, employ standardized metrics, use public datasets, and ensure external validation for generalizability.
