Computer-aided Diagnosis for Early-stage Lung Cancer Based on Longitudinal and Balanced Data

摘要

Objective: To facilitate the prediction of characteristics of SPNs in CT of lungs to diagnose early-stage lung cancer. Methods: All data was acquired from a cohort study. The synthetic minority over-sampling technique (SMOTE) was used to account for raw data in order to balance the original training data set. Input data, which included Curvelettransformation textural features, together with 3 patient demographic characteristics, and 9 morphological features were used to establish a SVM machine prediction model. Longitudinal data as the test data set was used to evaluate the classification performance of predicting early-stage lung cancer. Results: Using the SMOTE as a pre-processing procedure, the original training data was balanced with a ratio of malignant to benign cases of 1:1. Accuracy based on cross-evaluation (f=5) for the original unbalanced data and balanced data was 80% and 97%, respectively. Based on Curvelettransformation textural features and other features, the SVM prediction model had good classification performance for early-stage lung cancer, with an area under the curve of the SVMs of 0.949 (P<0.001). In addition to this, we found the textural feature (standard deviation) for benign cases had a higher change in the follow-up period than malignant cases. Conclusions: With textural features extracted from a Curvelet transformation and other parameters, a support vector machine prediction model sensitive to early-stage lung cancer can increase the rate of diagnosis for early-stage lung cancer. This scheme can be used as an auxiliary tool to differentiate between benign and malignant early-stage lung cancers in CT images.