| Pulmonary nodules are common lung diseases, and it is prone to misdiagnosis. The wide application of CT technology improves the efficiency of the doctor’s diagnosis at certain extent, In order to further improve the treatment effect, computer-aided diagnosis technology has gradually become a kind of important measure which can help doctors determine whether patients have pulmonary nodules or the nodules are benign or malignant.This paper studies segmentation of region of interest(ROI), feature extraction and pulmonary nodule detection based on CT images. The main content is:(1) Based on iterative threshold, this paper combines morphology and statistics of image size acquire pulmonary parenchyma with less trachea. With the utilizing of circular filter, false positive rate is lower and the candidate nodules are more complete.(2) Research on feature extraction method of gray image and then extracts gray scale, texture, shape features. Expounds the support vector machine(SVM) and using this algorithm for candidate nodules detection on the dataset of cooperation hospital and get the accuracy of 89.4%, the sensitivity of 90.9%, the specificity of 89.3%.(3) Research on convolutional neural network(CNN). This paper resizes ROI of the dataset of cooperation hospital to 34 x 34 as input map. Then design an 8-depth model to recognize the ROI is lung nodule or not. The accuracy, sensitivity and specificity rate reached 84.6%,82.5% and 86.7% respectively.(4) Based on all the research above, this paper designs and realizes a reliable, flexible, lung nodules compute aided diagnosis system on CT images. |
