Multi-Task Deep Learning Model for Automated Detection and Severity Grading of Lumbar Spinal Stenosis on MRI:...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Manual grading of lumbar spinal stenosis (LSS) faces challenges due to imaging variability and subjectivity.
• Automated assessment of LSS severity is needed for objective and reproducible clinical decision-making.

Purpose of the Study:

• To evaluate deep learning models (VGG19, ConvNeXt-Tiny, DINOv2) for feature extraction in automated LSS grading.
• To assess the generalizability of these features when combined with classical machine learning classifiers.

Main Methods:

• Pretrained deep learning models (VGG19, ConvNeXt-Tiny, DINOv2) extracted features from axial MRI images.
• Features were used to train Logistic Regression, SVM, and LightGBM classifiers.
• Models were internally trained and externally validated on data from the University of Phayao Hospital.

Main Results:

• VGG19 features with Logistic Regression achieved the highest external accuracy (0.9556) and F1-score (0.9558).
• External validation showed excellent discrimination (AUC 0.994-1.000).
• VGG19 and ConvNeXt-Tiny demonstrated good generalizability, while DINOv2 showed reduced performance.

Conclusions:

• Deep convolutional features, especially VGG19, combined with classical ML classifiers offer robust and generalizable LSS grading.
• CNN-based feature extraction remains effective for spinal imaging analysis.
• This automated approach provides a practical pathway for clinical decision support in LSS management.