Optimizing Vertebral Lesion Diagnosis: A CAD-Driven Machine Learning Classification System With Multi-Modal MRI Texture Analysis

The precise differentiation of vertebral lesions is a crucial aspect in clinical practice, given the spine pivotal role in human anatomy. The current study proposes an innovative Computer-Aided Diagnosis (CAD) system that builds a comprehensive assessment of vertebral lesions by integrating four distinct MRI modalities: T1-weighted, T2-weighted, Short tau inversion recovery (STIR), and Diffusion Tensor Imaging (DTI). This integration facilitates the acquisition of complementary information from each modality, allowing for a more precise diagnosis. The proposed system utilizes different voxel-based texture analysis techniques, including first-order statistics, the Gray-Level Co-occurrence Matrix (GLCM), and the Gray-Level Run Length Matrix (GLRLM), which help capture various vertebra texture characteristics such as homogeneity/inhomogeneity, morphology, and connectivity. The study included a total of 54 patients (25 benign, 29 malignant) with confirmed vertebral lesions. From each modality, a total of 69 texture markers were extracted, along with fractional anisotropy (FA) and mean diffusivity (MD) values from DTI. A comprehensive set of machine learning classifiers was utilized, coupled with Particle Swarm Optimization (PSO) techniques for feature selection and a Tree Parzen Estimator (TPE) for fine-tuning parameters. Upon combining all modalities, the AdaBoost classifier achieved 95.63% accuracy, 96.35% sensitivity, 99.38% specificity, 98.87% precision, 96.55% F1 score, and 96.67% balanced accuracy. These findings confirm the need for integrating different MRI modalities, especially in crucial locations such as the spine, to achieve a reliable and precise diagnosis.