The Automated Nodule Detection Algorithm Research Based On DoG Function

Recently, weather extremes like hazy weather appeared frequently, the air quality of many cities in China is continuous deteriorated, which directly increases the incidence of lung cancer. How to effectively improve curative effects has been a critical task in medical science. The clinical data of recent years pointed out that the diagnosis in early stage of lung cancer could obviously reduce the fatality rate. So far,the most effective way to diagnose lung cancer is using computed tomography(CT).The treatments of lung cancer is determined by observing the patient’s CT images.Thus, it is inevitable that treatments planed by different doctors hold different effects.In order to eliminate the difference of different treatments and help doctors to better diagnostic lung cancer, people tends to find helps from Computer-aided diagnosis system(CAD).Computer-aided diagnosis system uses CT imaging features and is combined with the knowledge of image processing and pattern recognition to help doctors to diagnose patients. The key part of CAD system is how to create an effectively algorithms. Generally, a CAD system is consisted of following steps: lung parenchyma field segmentation, nodule detection and segmentation, feature extraction and nodule diagnosis. Each step plays an important role in this system. To increase the accuracy of nodule detection and enhance the reliability of diagnosis, this paper fixes the attention on a nodule detection method, which is based on the Difference of Gaussian(Do G). Experimental results show that the proposed method is able toincrease the accuracy of nodule detection and the validity of lately feature extraction.The main contents of this paper are followings:1) Based on the features of lung field contour, image binarization and morphology methods are applied to segment lung parenchyma field. This method has the advantages ofless computation and fewer user participation.2) This paper proposes a method that locates nodule by Difference of Gaussian function to detect and segment nodules. This method overcomes the disadvantages of changing the shapes and sizes of nodules that are generated by filter method.3) Moreover, the proposed method takes advantage of the Hessian matrix to reduce false positive nodules. Experimental results showed that the proposed method has higher accuracy, less false positive rate and lower computation complexity comparing with the Laplace of Gaussian(LoG) function.4) 21 effective features, including gray level feature, configuration feature,geometric feature and textural feature are extracted. In this way it avoids the complicated computation and over fitting problem which caused by too many features.Furthermore, Support Vector Machine(SVM) is utilized to classify samples. To figure out the suitable parameter, this work has compared the results that under different situation. Experimental results show that the proposed method to detect and segment nodules achieves better performance than that of the filter based method.