Background: This study investigates the potential of classification and regression trees (CARTs) for the evaluation of thyroid lesions.
Methods: The study was performed on 521, histologically confirmed cytological specimens prepared via liquid based cytology. For each specimen, contextual and cellular morphology features were recorded by experienced cytopathologists, as described in everyday cytological practice and The Bethesda System (TBS); these features were subsequently used to construct two CART models, viz. CART-C for the prediction of the cytological diagnosis (according to TBS) and CART-H for the prediction of the histological diagnosis (hereby expressed as either benign or malignant).
Results: CART-C had no statistically significant performance from the cytologists' evaluations and CART-H had a very good predictive performance for the histological status.
Conclusion: CARTs provide a methodological framework capable for data mining and knowledge extraction. They created simple human understandable rules and classification algorithms that may assist cytopathologists towards decisions based on classification steps, each one linked with a specific risk and moreover by applying cytomorphological characteristics in hierarchical order according to their importance. The two CARTs may be a useful tool for the training of nonexperienced cytopathologists; moreover, they may act as ancillary methods to avoid misdiagnoses and assist quality assurance procedures in the everyday practice of the cytopathology laboratory.
Keywords: classification and regression trees; liquid based cytology; machine learning; morphology; thyroid cytopathology.
© 2018 Wiley Periodicals, Inc.