Related Experiment Video
Updated: Jun 23, 2025
Author Spotlight: Scope of LE-ULBD as a Safe, Effective, and Minimally Invasive Approach to Treat Lumbar Spinal Stenosis
Published on: February 9, 2024
Deep learning-based detection of lumbar spinal canal stenosis using convolutional neural networks
Hisataka Suzuki1, Terufumi Kokabu1, Katsuhisa Yamada2
1Department of Orthopaedic Surgery, Faculty of Medicine and Graduate School of Medicine, Hokkaido University, N15W7, Sapporo, Hokkaido 060-8638, Japan; Department of Orthopaedic Surgery, Eniwa Hospital, 2-1-1 Kogane Chuo, Eniwa, Hokkaido 061-1449, Japan.
This study developed a deep learning algorithm using convolutional neural networks (CNNs) to diagnose lumbar spinal canal stenosis (LSCS) from plain radiographs. The AI tool shows high accuracy, potentially reducing diagnostic delays for patients needing surgery.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Diagnostic Algorithms
- Radiographic Analysis of Spinal Disorders
Background:
- Lumbar spinal canal stenosis (LSCS) is a common degenerative condition in the elderly, often managed by non-specialists.
- Early diagnosis is crucial for effective surgical treatment, but Magnetic Resonance Imaging (MRI) may be unavailable or difficult to interpret.
- Convolutional Neural Networks (CNNs) offer advanced image recognition capabilities suitable for readily available radiographic facilities.
Purpose of the Study:
- To develop and validate a CNN-based algorithm for diagnosing surgically indicated LSCS using plain lumbar spine radiographs.
- To assess the algorithm's performance in identifying LSCS requiring surgical intervention.
Main Methods:
- Retrospective analysis of 600 lateral plain lumbar spine radiographs from 150 patients who underwent surgery for LSCS.
- Development of a CNN algorithm for binary classification (operative vs. nonoperative levels) and spinal canal area measurement.
- Internal validation using 5-fold cross-validation on 500 images, and external validation on 100 images from the same institution and 100 from other hospitals.
Main Results:
- Internal validation showed Area Under the Curve (AUC) of 0.85-0.89 and accuracy of 79-83% for classification.
- External validation achieved AUC of 0.90 and accuracy of 82%, with correlation coefficients for spinal canal area ranging from 0.56 to 0.69.
- Grad-CAM visualization highlighted feature importance in intervertebral joints and posterior intervertebral discs.
Conclusions:
- The developed CNN technology can automatically detect LSCS from plain radiographs.
- This facilitates diagnosis in facilities lacking MRI or specialist interpretation, potentially eliminating treatment delays.
- The algorithm shows promise for improving early detection and management of LSCS requiring surgery.
