Static cytometry and neural networks in the discrimination of lower urinary system lesions.

OBJECTIVES

To investigate the potential of morphometry and artificial intelligence tools for the discrimination of benign and malignant lower urinary system lesions.

METHODS

The study group included 50 cases of lithiasis, 61 cases of inflammation, 99 cases of benign prostatic hyperplasia, 5 cases of in situ carcinoma, 71 cases of grade I transitional cell carcinoma of the bladder (TCCB), and 184 cases of grade II and grade III TCCB. Images of voided urine smears stained by the Giemsa technique were analyzed by a custom image analysis system. The analysis gave a data set of features from 45,452 cells. A learning vector quantizer (LVQ)-type neural network (NN) was used to discriminate benign from malignant cells on the basis of the extracted morphometric and textural features. The data from 13,636 randomly selected cells were used as a training set and the data from the remaining 31,816 cells made up the test set. Similarly, in an attempt to discriminate at the patient level, 30% of the cases randomly selected were used to train an LVQ NN and the remaining 329 cases were used for the test.

RESULTS

The application of the LVQ NN enabled the correct classification of 95.42% of the benign cells and 86.75% of the malignant cells, giving an overall accuracy rate of 90.63%. At the patient level, the LVQ NN enabled the correct classification of 100% of benign cases and 95.6% of malignant cases, giving an overall accuracy rate of 97.57%.

CONCLUSIONS

NNs combined with image analysis offer useful information in the discrimination of benign and malignant cells and lesions of the lower urinary system.