Texture- and Object-related Automated Information Analysis in Histological Still Images of Various Organs

Objective: To create algorithms and application tools that can support routine diagnoses of various organs.

Materials: A generalized algorithm was developed that permits the evaluation of diagnosis-associated image features obtained from hematoxylin-eosin-stained histopathologic slides. The procedure was tested for screening of tumor tissue vs. tumor-free tissue in 1,442 cases of various organs. Tissue samples studied include colon, lung, breast, pleura, stomach and thyroid. The algorithm distinguishes between texture- and object-related parameters. Texture-based information-defined as gray value per pixel measure--is independent from any segmentation procedure. It results in recursive vectors derived from time series analysis and image features obtained by spatial dependent and independent transformations. Object-based features are defined as gray value per biologic object measured.

Results: The accuracy of automated crude classification was between 95% and 100% based upon a learning set of 10 cases per diagnosis class. Results were independent from the analyzed organ. The algorithm can also distinguish between benign and malignant tumors of colon, between epithelial mesothelioma and pleural carcinomatosis or between different common pulmonary carcinomas.

Conclusion: Our algorithm distinguishes accurately among crude histologic diagnoses of various organs. It is a promising technique that can assist tissue-based diagnosis and be expanded to virtual slide evaluation.