The Research Of Benign And Malignant Lung Nodules Auxiliary Diagnosis Based On CT Images

In recent years,lung cancer is the leading cause of the cancer-related deaths in humans worldwide.And early detection of cancer is the most promising way to enhance a patient9 s likelihood of survival,the survival rate is higher if the cancer is detected at early stages.Therefore,the identification of the potential malignant lung nodules is essential for the diagnosis of lung cancer.For the small nodules,due to the affect of the morphology imaging manifestation,the diagnosis of nodules for lung cancer is always the key point and the difficult point.However,the interpretation of CT images is challenging for the radiologists in common standard due to the anisotropic performances of nodules.And manual reading can be error-prone and the reader may miss the nodules.This thesis presents a nodule identification method in computed tomography images of lungs.It would be helpful for radiologists by offering initial screening or second opinions to classify lung nodules,and is capable of analyzing the large number of small nodules by CT scans.The small nodules can be classified into substantial nodules and subsolid nodules in the type.In consideration of the characteristics of the two types of nodules and the complex morphological features,this thesis studied the two types of nodules separately.For subsolid nodules(also known as ground glass nodules),this thesis used the patient's follow-up CT images to research the variation characteristics of the lung nodule size,and by predicting the multiplication rate to judge the conditions of lung nodules' growth and reliably identify the benign or malignant nodules.And based on the idea of the deep learning,with using the lung nodules' abstract image information,the substantial nodules' deep features have been extracted to ensure the accuracy of the nodule identification.In the lung cancer screening,the benign and malignant nodules can be classified through the nodule growth assessment by the registration and the subtraction between the follow-up computed tomography scans.During the registration,the volume of nodule regions in the floating image should be preserved,whereas the volume of other regions in the floating image should be aligned to those in the reference image.This thesis proposed an accurate and fast B-spline non-rigid registration method.And in order to take better advantage of the massive medical data,the thesis used unsupervised feature extraction methods.By following the basic idea of stack principle,this thesis proposed a deep sparse auto-encoder structure,named stacked auto-encoders to extract robust features for substantial lung nodules classification,and the method can be easily applied to any other image classification problems.