Learning vector quantizer in the investigation of thyroid lesions.

OBJECTIVE To investigate the capability of the learning vector quantizer (LVQ) in the discrimination of benign from malignant thyroid lesions. STUDY DESIGN The study was performed on May-Grunwald-Giemsa-stained smears taken by fine needle aspiration (FNA). Using a custom image analysis system, 25 features that describe the size, shape and texture of approximately 100 nuclei were measured from each case. Statistical features were extracted from each case, and a linear regression analysis was performed to detect the statistically significant features. The cases were distributed by category, as follows: 100 cases of goiter and follicular adenomas, 11 cases of follicular carcinoma, 35 cases of papillary carcinoma, 24 cases of oncocytic adenoma, 8 cases of oncocytic carcinoma and 20 cases of Hashimoto thyroiditis. About 30% of the cases from each class were used for training two LVQ classifiers. The remaining 139 cases, out of a total of 198, were used as the test set. A classifier was used to discriminate into four classes and a second into two classes. RESULTS The application of LVQ neural networks allows good discrimination between benign and malignant lesions (O.A. = 97.8). However, reliable discrimination of the cytologic types of the lesions was not obtained. CONCLUSION These results indicate that the use of neural networks combined with image morphometry may offer useful information on the potential for malignancy of thyroid lesions and may improve the diagnostic accuracy of FNA of the thyroid gland, especially in cases of follicular neoplasms classified as suspicious for malignancy and in cases of oncocytic tumors.