Technical Research On Computer Aided Diagnosis With 3D Facial Features

With the rapid development of computer science and medical research, Computer Aided Diagnosis (CAD) has become an important or even clinical standard tool for disease diagnosis. Now, the research of CAD is targeting the developing of Automated Computer Diagnosis (ACD).Genetic research shows that Genetic Syndromes (GS) are one of the main reasons, which lead to mental retardation. The syndromes will usually cause the change of facial morphology, and this makes it possible and necessary to diagnosis genetic syndromes with computer. The aim of this paper is to look into some important issues of CAD research and develop the system architecture with respect to this interdisciplinary topic.To extract facial features from facial 3D model is an important issue in CAD research. On considering of the application domain, the extraction needs to solve two problems: i) it is difficult to acquire frontal images, so that face orientation is required to be extracted on non-frontal 3D images; ii) normal extraction algorithms are not fit for the change of facial morphology, a solution is needed to deal with both normal and abnormal faces. To address the problems, this thesis presents a Nose Identification-based Feature Extraction (NIFE). The extraction algorithm starts by identifying the position of nose tip and its corresponding symmetry plane according to geometry characteristics. After that, 3D models are adjusted to be frontal based on the results of nose identification. Finally, facial regions are segmented by curvature parameters and the position to the nose tip, and then, feature points are extracted in their corresponding facial regions. NIFE is applicable on both normal and abnormal faces by using nose tip, which is stable in shape, as a reference point. The experiment results show that NIFE is fast and efficient to extract facial features on non-frontal 3D modelsIntelligent reasoning is the core of an ACD system. It must be ensured that the reasoning algorithm is accurate and has the ability of generalization. Because the experiment data is labeled, a supervised learning algorithm is a proper choice. With respect to the special requirements of ACD, chapter 3 introduces some typical machine learning algorithms. By analyzing these algorithms from the angles of the sample quality, algorithm accuracy and the incorporating of prior knowledge, the thesis chooses Support Vector Machine (SVM) as the engine of intelligent reasoning. The static learning theory and the implementation of multi-classification are discussed to show the validness of the choice.Incorporating prior knowledge is an efficient way to improve the accuracy of machine learning. To incorporate prior knowledge, we have to develop a proper diagnosis schema. Chapter 4 first gives a brief overview on state-of-art CAD/ACD schemas, and then arguing that current schemas are not able to meet the interdisciplinary requirement of CAD/ACD, and cannot accommodate the change of diagnosis knowledge. To address the problem, this thesis proposes an interdisciplinary-oriented ACD schema. The main idea is to incorporate medical expert in diagnosis schema by employing Natural Language Processing (NLP) so that prior knowledge can be incorporated as medical diagnosis instructions. To deal with the medical diagnosis instructions, production rules are proposed and implemented based on the discussion of NLP. At the end of this chapter, the thesis proposes an approach to incorporate prior knowledge. The incorporation acts on both sample selection and feature vector generation. With the interpreted prior knowledge, irrelevant samples are eliminated and feature vectors can be expanded with relevant dimensions. According to the experiment results, both of the operations are effective and the feasibility of the schema is proved.Because the experiment data comes from different institutions, the labels of training samples are not consistent. The inconsistent labeling brings a sub-separable problem. That is, samples with the same label may represent different sub classes. To address the problem, a Grouped Bagging SVM (GBSVM) is proposed. The GBSVM uses cluster algorithms to regroup the sub-separable samples and then construct sub classification SVM with an even sample selection procedure, which makes sure that the sub-separable samples with the same label will not be separated. On compared with ABSVM and normal SVM, GBSVM is more accurate and effective on the experimentdata.The last chapter of the thesis presents an ACD prototype system. The system isdesigned in a multi-layer manner with the theories and techniques applied. The entirediagnosis process is implemented including data capture and intelligent reasoning, andthe system is highly practical.