Machine Learning Predicts Decompression Levels for Lumbar Spinal Stenosis Using Canal Radiomic Features from Computed...

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning in Medicine

Background:

• Accurate preoperative identification of decompression levels is critical for multi-level lumbar spinal stenosis (LSS) surgery success.
• Machine learning (ML) classifiers were developed to predict decompression levels using computed tomography myelography (CTM) data.

Purpose of the Study:

• To develop and evaluate ML classifiers for predicting decompression levels in LSS patients based on CTM data.
• To identify key radiomic features predictive of decompression levels.

Main Methods:

• 1095 lumbar levels from 219 LSS patients were analyzed.
• Radiomic features were extracted from manually delineated bony spinal canals in CTM images.
• 72 ML classifiers were created by combining six feature selection methods with 12 ML algorithms.

Main Results:

• The embedding linear support vector (embeddingLSVC) was the optimal feature selection method.
• Top predictors included texture, intensity, and shape radiomic features.
• The EmbeddingLSVC_SVM classifier achieved ROC-AUC > 0.90 and PR-AUC > 0.80, demonstrating superior discrimination and calibration.

Conclusions:

• ML effectively extracted interpretable radiomic features from CTM for predicting LSS decompression levels.
• The EmbeddingLSVC_SVM classifier shows promise for assisting clinical surgical decision-making in LSS patients.