Detection and Classification of Thyroid Follicular Lesions Based on Nuclear Structure from Histopathology Images
JA Ozolek, W Wang, GK Rohde. University of Pittsburgh, Pittsburgh, PA; Carnegie Mellon University, Pittsburgh, PA
Background: Follicular adenoma (FA) and follicular carcinoma (FTC) are tedious challenges in surgical pathology due to lack of discriminatory cytological and microarchitectural features. Limitations of the diagnostic algorithm include time consuming tissue processing and microscopic evaluation. The aim was to develop an automated image analysis technique that could classify these lesions with 100% accuracy from routinely processed tissue using nuclear structure.
Design: Cases included 5 FA and 5 FTC resections. Sections were stained using Feulgen technique. Nuclei were segmented using random field graph cut and efficient level set active contour algorithms to yield 871 NL, 489 FA, and 703 FTC nuclei. 125 features were extracted from each nucleus. Four different classifiers (Mahalanobis distance nearest neighbors and support vector machine with different kernels) and voting strategy were used. Unique chromatin patterns were identified in feature space by finding nuclei near to each other and most distant from nuclei in other classes.
Results: These methods automatically classify the data with 100% accuracy after blind cross validation using at most 43 nuclei randomly selected from each patient.