Objective: To investigate the potential of back propagation (BP) neural networks (NNs) in the discrimination of benign from malignant thyroid lesions.
Study design: The study was performed on May-Grünwald-Giemsa-stained smears obtained by fine needle aspiration (FNA). Using a custom image analysis system, 26 features that describe the size, shape and texture of the nucleus were measured from each cell. The cases were distributed according to categories, as follows: 25 cases of goiter and follicular adenomas, 1 case of follicular carcinoma, 12 cases of papillary carcinoma, 6 cases of oncocytic adenoma, 3 cases of oncocytic carcinoma and 4 cases of Hashimoto thyroiditis. From each case about 100 nuclei were measured; they formed a pool of 13,850 feature vectors. Out of this pool, 2,770 vectors were randomly selected to form the training set, and the remaining 11,080 vectors formed the test set.
Results: The application of a BP NN on the nuclear measurements permitted correct classification of 90.61% nuclei. Classification at the patient level was performed using a hypothesis test for proportion and two different hypothesis values, one equal to the overall accuracy of the NN and one equal to 50%. The second method permitted correct classification of 98% of patients.
Conclusion: These results indicate that the use of NNs combined with image morphometry and statistical techniques may offer useful information on the potential malignancy of thyroid cells and may improve the diagnostic accuracy of FNA of the thyroid gland, especially in cases classified as suspicious for malignancy.