Frequential versus spatial colour textons for breast TMA classification.

Area of Science:

• Digital pathology
• Computational pathology
• Biomedical image analysis

Background:

• Digital pathology generates vast whole slide (WSI) and tissue microarray images (TMA) for cancer research.
• Extracting comprehensive information from heterogeneous tissue data is a significant challenge.

Purpose of the Study:

• To identify optimal features for classifying breast TMA images.
• To differentiate between stroma, adipose tissue, benign structures, and carcinomas (ductal and lobular).

Main Methods:

• Exhaustive assessment of textons and color for automatic breast TMA classification.
• Extraction and comparison of frequential and spatial texton maps across eight color models.
• Characterization using Haralick statistical descriptors from texton maps (241x8 features).
• Feature selection using linear discriminant analysis, correlation, and sequential forward search.
• Comparison of six classifiers, focusing on Fisher, Bagging Trees, and AdaBoost.

Main Results:

• Spatial texton maps combined with specific color models (Hb, Luv, SCT) provide efficient breast TMA representation.
• AdaBoost classifier demonstrated superior performance across all color model combinations.
• Achieved 98.1% accuracy and 96.2% precision on 628 TMA images using 316 features.

Conclusions:

• The combination of spatial texton maps and selected color models is highly effective for breast TMA classification.
• The AdaBoost classifier, with 316 spatial texton features, offers a robust solution for automated breast cancer detection in digital pathology.