Local Binary and Ternary Patterns Based Quantitative Texture Analysis for Assessment of IDH Genotype in Gliomas on Multi-modal MRI.

Radiomics based multivariate models are potentially valuable prognostic tool in IDH phenotyping in high-grade gliomas. Radiomics generally involves a set of the histogram and standard texture features based on co-occurrence matrix, run-length matrix, size zone matrix, gray tone matrix, and gray level dependence matrix as described by the Imaging Biomarker Standardization Initiative (IBSI). In this work, we introduce 3D local binary patterns (3D LBP) and local ternary patterns (3D LTP) based histogram features in addition to standard radiomics for capturing subtle phenotypic differences to characterize IDH genotype in high-grade gliomas. These textures are rotationally invariant and are robust to image noise as well as can capture the underlying local differences in the tissue architecture. On a dataset of 64 patients with high-grade glioma scanned at a single institution, we illustrate that LBP and LTP features perform at par with radiomics individually and when combined, could facilitate highest testing accuracy of 84.62% (AUROC: 0.78, sensitivity: 0.83 specificity: 0.86, f1-score: 0.83) using random forest classification. Clinical explainability is achieved via feature ranking and selection where the top 5 features are based on LBP and LTP histogram. Overall, our results illustrate that 3D LBP and LTP histogram features are crucial in creating tumor phenotypic signatures and could be included in the standard radiomics pipeline.