Radiomic Machine Learning Classifiers in Spine Bone Tumors: A Multi-Software, Multi-Scanner Study.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Spinal lesion differential diagnosis is challenging with MRI alone.
• Radiomics and machine learning (ML) offer potential solutions despite a lack of standardized pipelines.
• Previous studies show ML's utility in medical image analysis.

Purpose of the Study:

• To evaluate the diagnostic performance of ML in spinal lesion differential diagnosis.
• To compare ML performance using radiomic data extracted by different software.
• To assess the impact of feature selection and ML algorithms on diagnostic accuracy.

Main Methods:

• Retrospective analysis of 146 patients with vertebral lesions undergoing MRI.
• Radiomic features extracted using 3D Slicer heterogeneityCAD (hCAD) and PyRadiomics.
• Two-label (benign/malignant) and three-label (benign/primary malignant/metastases) classification models were trained and tested on internal and external cohorts.

Main Results:

• ML achieved 94% accuracy for 2-label classification using hCAD data in the internal test cohort.
• Accuracy for 2-label classification was 86% in the external test cohort.
• For 3-label classification, PyRadiomics data yielded 80% accuracy internally and 69% externally.

Conclusions:

• MRI radiomics combined with ML demonstrates potential for spinal lesion assessment.
• Robust preprocessing enhances consistency across heterogeneous scanner and protocol conditions.
• This approach may aid in improving the accuracy of spinal lesion differential diagnosis.