Application of the learning vector quantizer to the classification of breast lesions.

Objective: To investigate the potential of the learning vector quantization (LVQ) neural network for the discrimination of benign from malignant breast lesions.

Study Design: Using a custom image analysis system on Giemsa-stained smears, 25 parameters describing the size, shape and texture of the cell nucleus were measured. Three thousand nuclei from a total of 9,356 were selected as a training set for the neural network, and the whole data set was used for testing. An additional 238 cells from 16 cases without final cytologic diagnoses were evaluated by the system. The total number of cells (9,594) was collected from 100 patients (68 carcinomas and 32 benign lesions).

Results: Cytologic examination of the cases gave two false positive and two false negative results. However, in eight cases of ductal breast carcinoma and in eight cases of benign lesions, histologic confirmation was necessary in order to confirm the cytologic diagnosis. Application of the LVQ permitted correct classification of 87.41% of the cells. Classification at the patient level by using a hypothesis test for proportion with a hypothesis value equal to 50% permitted the correct diagnosis in 98% of patients.

Conclusion: These results indicate that the use of neural networks combined with image morphometry and statistical techniques may offer useful information about the potential for malignancy, improving the diagnostic accuracy of fine needle aspiration of breast lesions.