Tree-based Methods for Characterizing Tumor Density Heterogeneity.

Area of Science:

• Radiology
• Medical Imaging
• Computational Pathology

Background:

• Solid lesions pose diagnostic challenges due to diverse tissue environments.
• Cancer radiomics utilizes quantifiable feature sets from images to describe lesion morphology and texture.
• Existing methods for tissue density heterogeneity characterization rely on empirical distributions and grey-level co-occurrence statistics.

Purpose of the Study:

• To propose a novel decision-tree based approach for quantifying tissue density heterogeneity in lesions.
• To develop new metrics that add novel information to the radiomics feature space.
• To improve the quantitative characterization of heterogeneous density distributions in CT images.

Main Methods:

• A novel decision-tree based approach is proposed.
• Quantification of tissue density heterogeneity is achieved through tree-structured dissimilarity metrics.
• Metrics are computed using least common ancestor trees under repeated pixel re-sampling, based on Galton-Watson tree statistics.

Main Results:

• The proposed methodology produces metrics minimally correlated with existing radiomics features.
• The new metrics add valuable information to the feature space for enhanced lesion characterization.
• Integration with existing features improved classification accuracy for adrenal lesions: malignant vs. benign (AUC = 0.78), functioning vs. non-functioning (AUC = 0.93), and calcified vs. non-calcified (AUC = 1).

Conclusions:

• The novel decision-tree based approach effectively quantifies tissue density heterogeneity.
• The developed metrics enhance the radiomics feature space, offering new diagnostic information.
• This method significantly improves the diagnostic classification of adrenal lesions.