基于卷积神经网络检测肺结节

Detection of pulmonary nodules based on conventional neural networks

Objective Major challenges in the current automatic detection of lung nodules from chest CT images are to improve the sensitivity and to reduce the false positive rate. A new scheme based on convolutional neural network was proposed in this study. Methods The method applied an automatic anatomy recognition (AAR) methodology based on fuzzy modeling ideas and an iterative relative fuzzy connectedness (IRFC) delineation algorithm for the segmentation of lung parenchyma in CT images. The segmented lung image was inputted into the conventional neural networks for feature extraction of pulmonary nodules. The network adopted position-sensitive score maps to express the location information of lung nodules. Results This method could obtain accurate segmentation of the lung parenchyma in the data set of Tianchi Medical AI Contest, and the accuracy, sensitivity, specificity and false-positive rate of lung nodules detected was 95.60%, 95.24%, 95.97% and 4.03%, respectively. Conclusion Detection of pulmonary nodules based on convolutional neural networks has high accuracy and efficiency, and good robustness.