OBJECTIVE: To investigate capability of combination of learning vector quantizer (LVQ) neural networks (NNs) in discrimination of benign from malignant thyroid lesions. STUDY DESIGN: The study included 335 liquid-based cytology, fine needle aspiration (FNA), Papanicolaou-stained specimens. All cases were compared to the histologic diagnosis. Features describing size, shape, and texture of -100 nuclei per case were extracted from cytologic images using a custom image analysis system. These features were used to classify each nucleus by LVQ type NNs. The nucleus classification results were used to classify individual lesions with a second LVQ NN. Cases were distributed by histologic diagnosis. Data from -50% from each category were used for training LVQ classifiers. Remaining data were used to test classifier performance. The system was used to discriminate to individual cellular level and individual patient level between benign and malignant nuclei. RESULTS: Application of the proposed algorithm combining two LVQ NNs allows discrimination between benign and malignant cell nuclei and lesions. CONCLUSION: Results indicate that use of NNs, combined with image morphometry, can provide information on thyroid lesion malignancy potential. The system could improve FNA diagnostic accuracy of the thyroid gland, especially in follicular neoplasms suspicious for malignancy and in Hürthle cell tumors.
OBJECTIVE: To investigate capability of combination of learning vector quantizer (LVQ) neural networks (NNs) in discrimination of benign from malignant thyroid lesions. STUDY DESIGN: The study included 335 liquid-based cytology, fine needle aspiration (FNA), Papanicolaou-stained specimens. All cases were compared to the histologic diagnosis. Features describing size, shape, and texture of -100 nuclei per case were extracted from cytologic images using a custom image analysis system. These features were used to classify each nucleus by LVQ type NNs. The nucleus classification results were used to classify individual lesions with a second LVQ NN. Cases were distributed by histologic diagnosis. Data from -50% from each category were used for training LVQ classifiers. Remaining data were used to test classifier performance. The system was used to discriminate to individual cellular level and individual patient level between benign and malignant nuclei. RESULTS: Application of the proposed algorithm combining two LVQ NNs allows discrimination between benign and malignant cell nuclei and lesions. CONCLUSION: Results indicate that use of NNs, combined with image morphometry, can provide information on thyroid lesion malignancy potential. The system could improve FNA diagnostic accuracy of the thyroid gland, especially in follicular neoplasms suspicious for malignancy and in Hürthle cell tumors.