Lung Nodule Detection Based On CT Images

Lung cancer is becoming one of the biggest threats to human health,whoes morbidity and mortality have increased year by year.The most effective measures are ”early detection,early diagnosis and early treatment”.For the early manifestation of lung cancer is pulmonary nodules,the detection of nodules plays an important role in the early diagnosis of lung cancer.At present,lung cancer is mainly diagnosed by scanning CT images with the naked eye.With the development of medical technology,the resolution of CT scan has been improved a lot,leading to the growing amount of data and extra burden of doctors.Computer aided diagnosis(CAD)system is designed to assist doctors to diagnose effectively.Lung cancer CAD system usually includes four main functions: lung segmentation,lung nodule detection,lung nodule segmentation and identification of benign and malignant lung nodules.The pulmonary nodule detection can be divided into two parts: nodule candidate extraction and false-positive reduction.This thesis mainly deals with the first two functions,the main contributions are detailed as follows:1.A threshold-based three-dimensional automatic lung segmentation algorithm is proposed,which includes four steps: lung extraction,trachea and main bronchi elimination,left and right lung separation and lung border revise.The algorithm is mainly improved in two aspects: for the current trachea segmentation algorithms require experts to select the initial seed points,a three-dimensional automatic region growing method is proposed in this paper;Aiming at the problem that the parameters of rolling ball method are difficult to adjust,an adaptive lung boundary revise method is proposed.2.A nodule candidate extraction algorithm based on contours evolution is proposed.At present,most of the pulmonary nodule extraction algorithms are designed for fixed types of pulmonary nodules.Based on the fact that different types of pulmonary nodules are locally or globally showing a certain degree of circular nature,and contours have a good ability to describe the local area shape,we design a nodule candidate extraction algorithm which is not affected by the gray value and shows good robustness to noise.3.A false-positive reduction algorithm based on recurrent convolutional neural network is proposed.Traditional false-positive reduction algorithm relies much on feature engineering to train a classifier.In addition,the association between the CT layers provides key information for the diagnosis of pulmonary nodules which the traditional algorithm is difficult to catch.In this paper,the proposed false-positive reduction algorithm combines the strong feature extraction ability of convolution network and the sequence learning ability of the LSTM network(Long Short-term Memory),which achieves good experimental results.