Establishment Of Mathematical Model Of Human Lung Airway Tree Based On CT Volume Data

Lung disease is the major threat to human health, and the construction of Computer-Aided Diagnosis (CAD) system of lung disease has become a focus of research. At present, there are two different approaches to implement CAD system: the traditional Abnormality Dependent-Approach (ADA) and the novel Normality-Dependent Approach (NDA). ADA focused on the image region with obvious features of a specific disease, thus cannot deal with the common clinical cases that multiple diseases might exist simultaneously. NDA is more in line with the observation method and diagnostic thinking of professional doctors, that is, firstly identify and exclude the normal image area, then analysis the rest probably abnormal region more carefully and finely. In NDA, a variety of disease information can be extracted, not just a specific disease. Therefore NDA has become the development trend of the future lung disease CAD systems.Following the basic theories and principles of NDA, a novel technical route to construct the lung disease CAD system, in which the basic work is to establish a set of healthy lung parameters in computer, was proposed. Main attention was focused on parameters of lung airway, since airway disease are part of the most serious lung diseases and our work will show that airway parameters reflect the lung situation to a certain extent. Finally, a mathematical model with rich lung airway parameters was established and five steps were taken to achieve this goal. In each step, we conducted a deep research and achieved some initial results.Step 1:Lung Segmentation. The results of lung segmentation are the basis for subsequent airway segmentation. In order to solve the difficulty of fuzzy boundary, complex background, and uneven local part in the segmentation of medical images, a united active contours model based on relative fuzzy connectedness was proposed and proved the applicability for lung images in both theoretical aspect and experiment aspect. The proposed method achieved good results in images with multi-object and complexity background and finally got the result of lung segmentation with completeness and rightness.Step 2:Airway Segmentation. Airway segmentation is the basis of the model establishment and directly determines the performance of the model. Taking the segmentation results of ordinary algorithms as the seeds, the general framework of airway segmentation was improved segmentation module, evaluation module and etc. Targeted strategies were proposed to deal with the two problems of leakage and thin airway extracting. Finally,104 airway segments in 10 levels were extracted, which equaled to 45% of the number of segments segmented by manual segmentation, and the whole airway information above segmental branch were completely preserved.Step 3:Airway Skeletonization. Extracting single-pixel-width central skeleton of airway tree is the essential part for parameters measurement, and almost all the definitions and measurement methods of the structural parameters are based on the skeleton. After comprehensive analysis for the four classes of commonly used skeletonization algorithm in the specific needs of airway tree, the hierarchical general potential field method was selected. The entire skeletonization process was divided into three completeness-increasing levels:extracting the core skeleton (level 1), adding the thin airway branch (level 2) and connecting peripheral points (level 3), and the result of level 3 is the final skeletonization result. The proposed method achieved superiority of completeness compared to other methods.Step 4:Parameters Measurement. This is the crucial step directly related with the model establishment and the measured parameters set is the data part of the mathematical model. The whole airway tree was anatomized into four accuracies:by all, by level, by segment and by slice; and one category of parameters were defined and measured for each accuracy. There are 4 kinds of overall parameters,4 kinds of level parameters (for 10 levels),5 kinds of segment parameters (for 104 segments),5 kinds of slice parameters (for 1916 slices) and therefore totally 10144 parameters for the whole airway tree. Some of the parameters are compared with the facts of medical imaging and the real value of anatomy, and the comparing results show that these parameters'value are correct and believable.Step 5:Model Establishment. This is the ultimate goal of this dissertation. The established model consists of two parts:data and operation. All parameter data of the airway tree was stored in a tree structure with five levels. Each level of the tree consists of a kind of structures, which were especially for overall, level, segment, slice and voxel data. Operations on data, which included parameter computing, storing, loading, querying, calling and etc, were the second part of the model. Codes and pointers were preserved in structures for queries and calls. Three solutions for practical applications were given on the basis of model data and operation, as the reference for subsequent construction of lung disease CAD system.In summary, the established mathematical model can be used for constructing lung disease CAD system in the next step since there are rich and correct parameters of lung airway tree in it.The author appreciates National Natural Science Foundation of China (60771007) and Graduate Students Innovation Foundation of Chinese Academy of Science and Technology (2008 Year).