Machine Learning Classification of Spinal Lesions: Compared Accuracy of Texture Parameters Extracted by Different Software

Purpose: To compare the accuracy of machine learning (ML) algorithms for the classification of spinal lesions based on texture analysis (TA) parameters extracted by different software from unenhanced magnetic resonance imaging (MRI).

Methods and Materials: We retrospectively enrolled 146 patients with 146 spinal lesions (49 benign, 57 metastatic, and 40 primary malignant lesions) imaged using MRI. Of them, 117 were subsequently confirmed histopathologically after surgery; 29 benign lesions were confirmed by follow-up. Patients were randomly divided into training (n = 100) and test groups (n = 46), respectively, for classification model development and testing. Lesions were manually segmented on T1-weighted and T2-weighted images by drawing a bidimensional polygonal region of interest. These were used for first order and texture feature extraction on two types of software, 3D-Slicer heterogeneity CAD module (hCAD) and PyRadiomics. For each of them, different data subsets, obtained by four feature selection methods, were analyzed by nine ML classification algorithms to evaluate their accuracy in identifying benign versus malignant lesions and benign versus primary malignant versus metastatic lesions.

Results: In the test group, a random forest (RF) algorithm correctly classified 89% of lesions as benign or malignant, based on hCAD TA; a Support Vector Machine could achieve an accuracy of 87% from PyRadiomics TA. For the classification of benign, primary malignant, and metastatic lesions, RF models accurately classified 70% of lesions for both types of TA software.

Conclusion: ML algorithms show good accuracy in spinal lesion classification based on noncontrast MRI examinations. Furthermore, feature extraction performed using different software has shown consistent results at subsequent ML analysis.

No conflict of interest has been declared by the author(s).