Pantazopoulos D, Karakitsos P, Iokim-Liossi A, Pouliakis A, Botsoli-Stergiou E, Dimopoulos C
Department of Urology, University of Athens, Greece.
J Urol. 1998 May;159(5):1619-23. doi: 10.1097/00005392-199805000-00057.
We investigated the potential value of morphometry and artificial intelligence tools to discriminate between benign and malignant lower urinary tract lesions.
The lesions included lithiasis in 50 cases, inflammation in 61, benign prostatic hyperplasia in 99, carcinoma in situ in 5, and grade I and grades II and III transitional cell carcinoma of the bladder in 71 and 184, respectively. Images of routine processed voided urine smears stained by the Giemsa technique were analyzed using a custom image analysis system, providing a data set of 45,452 cells. A neural net model of the back propagation type was used to discriminate benign from malignant cells based on the extracted morphometric and textural features. Data from 13,636 randomly selected cells (30% of the total data) were used as a training set and the data from the remaining 31,816 cells comprised the test set. In a similar attempt to discriminate at the patient level data on 30% of those randomly selected were used to train a back propagation neural net and data on the remaining 329 were used for testing.
Application of the back propagation neural net enabled the correct classification of 95.34% of benign and 86.71% of malignant cells with overall 90.57% accuracy. At the patient level the back propagation neural net enabled the correct classification of 100% of those with benign and 94.51% of those with malignant disease with overall 96.96% accuracy.
The use of neural nets and image morphometry may increase the speed of cytological diagnosis and the diagnostic accuracy of voided urine cytology.
我们研究了形态测量学和人工智能工具在鉴别下尿路良性和恶性病变方面的潜在价值。
病变包括50例结石、61例炎症、99例良性前列腺增生、5例原位癌以及分别为71例和184例的膀胱I级、II级和III级移行细胞癌。使用定制图像分析系统对经吉姆萨技术染色的常规处理的排尿后尿液涂片图像进行分析，提供了一个包含45452个细胞的数据集。基于提取的形态测量和纹理特征，使用反向传播类型的神经网络模型来区分良性和恶性细胞。从13636个随机选择的细胞（占总数据的30%）中获取的数据用作训练集，其余31816个细胞的数据构成测试集。在类似的患者水平鉴别尝试中，随机选择的30%的数据用于训练反向传播神经网络，其余329例的数据用于测试。
应用反向传播神经网络能够正确分类95.34%的良性细胞和86.71%的恶性细胞，总体准确率为90.57%。在患者水平，反向传播神经网络能够正确分类100%的良性患者和94.51%的恶性疾病患者，总体准确率为96.96%。
使用神经网络和图像形态测量学可能会提高细胞学诊断的速度以及排尿后尿液细胞学的诊断准确性。
