Comparing neural networks in the discrimination of benign from malignant lower urinary tract lesions

To compare the performance of two different neural networks (NNs) in the discrimination of benign and malignant lower urinary tract lesions.

A group of patients was evaluated, comprising 50 cases of lithiasis, 61 of inflammation, 99 of benign prostatic hyperplasia (BPH), five of in situ carcinoma, 71 of grade I transitional cell carcinoma of the bladder (TCCB), and 184 of grade II and grade III TCCB. Images of routinely processed voided urine smears were stained using the Giemsa technique and analysed using an image-analysis system, providing a dataset of 45452 cells. Two NN models of the back propagation (BP) and learning vector quantizer (LVQ) type were used to discriminate benign from malignant cells and lesions, based on morphometric and textural features. The data from 13636 randomly selected cells (30% of the total data) were used as a training set and data from the remaining 31816 cells comprised the test set. Similarly, in an attempt to discriminate patients, 30% of the cases, selected randomly, were used to train a BP and an LVQ NN, with the remaining 329 cases used for the test set. The data used for training and testing were the same for the two kinds of classifiers.

The two NNs gave similar results, with an overall accuracy of discrimination of approximately 90.5% at the cellular level and of approximately 97% for individual patients. There were no statistically significant differences between the two NNs at the cellular or patient level.

The use of NNs and image morphometry could increase the diagnostic accuracy of voided urine cytology; despite the different nature of the two classifiers, the results obtained were very similar.