| Sumario: | Tumor heterogeneity is a significant cause of treatment resistance and poor clinical outcomes, and the medical images used to diagnose and treat patients are routinely interpreted in a largely qualitative manner and may not reflect subtle patterns of heterogeneity early in disease progression. This study is aimed at proposing an integrated mathematical modeling and image processing framework for quantification of intra-tumoral heterogeneity and early detection using medical imaging. Paired magnetic resonance and computed tomography brain tumor images along with the expert-annotated segmentation masks were analyzed. Following the data ingestion and quality control process, images were standardized and modality specific preprocessing pipelines were followed. Tumor regions of interest were extracted in order to calculate morphological descriptors, first order radiomic features, texture measures, and multi-scale entropy maps. A composite heterogeneity index was developed from the combination of entropy statistics, dispersion measures, and spatial autocorrelation to describe both the variation of intensity and the organization of space in tumors. Slices-level heterogeneity measures were aggregated to patients level and an interpretable L1 regularized logistic regression model was applied for the early vs. advanced heterogeneity discrimination under patient-wise cross-validation. The proposed heterogeneity-informed model showed stable performance and obvious contributions of spatial heterogeneity components. The results show that mathematically derived heterogeneity indices can be robust and interpretable imaging biomarkers to aid in the early characterization of tumor complexity and provide motivation for their validation in larger, clinically annotated cohorts.
|
|---|