Preliminary application of neural network in differentiating benign from malignant solitary pulmonary nodule on HRCT

In this study, the performance of artificial neural network (NN) in diagnosis of solitary pulmonary nodule (SPN) on HRCT images was evaluated. One hundred and forty-five cases of SPN, including 86 cases of pulmonary carcinoma, 18 cases of tuberculoma, 29 cases of inflammatory nodule and 12 cases of benign tumor, were collected, which were all confirmed by pathology or biopsy and over-two-year clinical treatment. Five clinical parameters and 10 radiological characteristics were observed and quantified for qualitative characteristics. About 70 percent of all cases (up to 103 cases) were selected randomly to form training samples set, on which BP neural network and Logistic regression model were built. The total consistent rate 98.6% of BP NN was greater than that of Logistic model, which is 88.3% (P=0.0007). Areas under ROC curve were 0.997+/- 0.004 and 0.959+/-0.016 respectively, and the difference between the two was significant statistically (P=0.009). NN showed high performance in diagnosis of SPN on HRCT images. It was worthy of further study.