Automated Severity Grading of Lumbar Spinal Stenosis on Sagittal MRI Using a Two-Stage Deep Learning Framework with External Clinical Validation.
Authors
Affiliations (7)
Affiliations (7)
- Universidade Politécnica de Setúbal, Escola Superior de Tecnologia de Setúbal, DICE - ALGORITMI Lab, Campus da UNIPS, Estefanilha, Setúbal, 2910-761, Portugal. [email protected].
- Neurosurgery Department, Hospital da Luz Setúbal, EN 10, km 37, Setúbal, 2900-722, Portugal.
- Neurogenetics of Locomotion, iNOVA4Health, NOVA Medical School/Faculdade de Ciências Médicas, NMS|FCM, Universidade NOVA de Lisboa, Lisbon, 1169-056, Portugal.
- Imagiology Department, Hospital da Luz, Av Lusiada 100, Lisbon, 1500-650, Portugal.
- Universidade Tecnológica Federal do Paraná (UTFPR) - Campus Medianeira, Av. Brasil, 4232, Medianeira, 85884-000, Brazil.
- Universidade Politécnica de Setúbal, Escola Superior de Tecnologia de Setúbal, DICE - ALGORITMI Lab, Campus da UNIPS, Estefanilha, Setúbal, 2910-761, Portugal.
- ALGORITMI Research Center/CCPM, Campus de Azurém, Guimarães, 4800-058, Portugal.
Abstract
Lumbar spinal stenosis (LSS) is a prevalent degenerative condition and a major cause of disability in aging populations. Magnetic resonance imaging (MRI) is the preferred modality for assessment; however, severity grading remains inconsistent across readers and institutions, limiting reproducibility and complicating clinical decision-making. This study presents LSS-Net, a two-stage deep learning framework for automated grading of central canal stenosis using sagittal T2-weighted MRI. The first stage performs vertebral localization using a single-stage object detector, enabling standardized extraction of intervertebral regions of interest. The second stage applies a Swin Transformer-based classifier to predict stenosis severity (Normal/Mild, Moderate, Severe). The framework was developed using a multi-center dataset of 1,975 studies and evaluated on a held-out subset. Vertebral localization achieved 97.9% F1-score, while severity classification reached a weighted F1-score of 77.1%. Performance was highest for Normal/Mild and Severe classes, with reduced accuracy observed in Moderate cases, reflecting known diagnostic ambiguity. External clinical validation was conducted on an independent dataset using two expert readers. The model demonstrated substantial agreement with a neurosurgeon (accuracy 75.1%, <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>κ</mi> <mo>=</mo> <mn>0.73</mn></mrow> </math> ) and moderate agreement with a neuroradiologist (accuracy 51.0%, <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>κ</mi> <mo>=</mo> <mn>0.56</mn></mrow> </math> ), with inter-expert agreement in a comparable range. These findings suggest that LSS-Net may support more consistent and reproducible LSS grading. However, performance limitations in borderline cases and the restricted use of single-slice sagittal input highlight the need for further validation and extension to multi-sequence and multi-compartment analysis before clinical deployment.