Computational Intelligence And Machine Learning Based Lymph Node Detection And Lymph Node Metastasis Diagnosis In Gastric Cancer

Gastric cancer is one of the second leading causes for cancer-related death in the world. Lymph Node Metastasis(LNM) is one of the most important prognostic factors regarding long-term survival. Doctors need to know the situation of lymph nodes as much as possible. Currently, lymph nodes are detected mainly according to imaging, and multiple indicators can be extracted. Then the useful indicators can be selected by using some indicator selection methods, and lymph node metastasis is diagnosed. Therefore, the problem can be divided into two sub-problems. The thesis is supported in part by the National Natural Science Foundation of China, and so on. Based on gastric cancer imaging, artificial intelligence, pattern recognition, image processing, machine learning and evolutionary computation, the problem has been studied and the corresponding solutions have been proposed. The innovative results obtained in this dissertation include:(1) Lymph nodes are detected based on imaging. According to the clinical domain knowledge, lymph nodes are usually distributed within the main fatty tissues which should be extracted accurately around the stomach. As the structures and textures are complicated and fatty tissues vary from patient to patient, automatic extraction is still a challenging task. Manual extraction is time-consuming and frequently not applicable in clinical routine. Consequently, it appears more realistic to develop methods enabling to interactively guide extracting process in clinical use. Therefore, an Object Information based Interactive Segmentation(OIIS) method is proposed. Different from the most existing methods, OIIS only needs to input markers on a small part of the fatty tissue, and background markers are not required. Meanwhile, to extend OIIS to 3-dimensional image, an Object Information based Interactive 3-dimensional Segmentation(OII3S) method is proposed. In OII3 S, radiologists just need to input markers in one or several slices, and fatty tissues in all slices are extracted. Experimental results and comparative studies demonstrate both of the proposed two methods are effective for fatty tissue extraction and can be fully used in the following lymph node detection.(2) To detect lymph nodes, according to the radiologists’ experience, a hierarchical model for lymph node detection is proposed. It consists of four layers: original image layer, fatty tissue layer, lymph node candidate layer and lymph node layer. To detect lymph nodes easily, fatty tissue regions should be extracted in the second layer. As there are vessels and other regions which are similar with lymph nodes, it is difficult to distinguish them directly. Hence, lymph node candidate is detected in the third layer. Finally, lymph nodes are detected. For fatty tissue layer, OIIS and OII3 S are employed. For lymph node candidate detection, a new Sparse Dynamic Ensemble Selection(SDES) method is proposed. Then an Evidential Reasoning(ER) based method is proposed for lymph node detection. The experimental results demonstrate that the proposed model is a valid approach for lymph node detection.(3) After lymph nodes are detected, multiple indicators should be extracted for LNM diagnosis. However, if fewer indicators are used, doctor’s workload can be reduced greatly. Moreover, although doctors can extract many indicators, they do not know which indicators are more useful for LNM diagnosis. What’s more, some indicators may have negative effect for diagnosing result. As such, it is important for doctors to select useful indicators. To solve this problem, a Correlation based Clonal Selection Algorithm(CCSA) for indicator selection is proposed. In CCSA, indicator selection is considered as the combinatorial optimization problem. The experimental results demonstrate that CCSA can obtain better performance, while it selects fewer indicators. Moreover, the selected indicators are meaningful for LNM diagnosis.(4) The literature study shows that there are two way for LNM diagnosis:(1) lymph node indicator based LNM diagnosis;(2) Tumor and lymph node indicator based LNM diagnosis. Currently, the available methods for LNM diagnosis are block-box modelling approaches in nature and their internal structures are not directly linked to the reasoning logic or process, which makes it difficult for doctors to know how important each medical attribute is regarding prediction results.To fully utilize the clinical domain knowledge, based on the lymph node indicator, a Bi-level Belief Rule Based(BBRB) model is proposed. BBRB consists of two layers, which are modeled by a single-layer BRB. As manually constructed belief rules may not be accurate, there is a need to train BBRB. Due to this issue, a novel Clonal Selection Algorithm(CSA) for training BRB system is proposed. Compared to conventional methods, new CSA is capable of improving performance significantly.On the other hand, based on tumor and lymph node indicator, a Cooperative Belief Rule Based(CBRB) model is proposed, which consists of two independent BRBs. One is used for modelling lymph node indicators, while the other one is used for modelling tumor indicators. Moreover, as manually constructed belief rules in CBRB may not be accurate, it is necessary to train the proposed CDSS. As CBRB is a cooperative model, a corresponding new Cooperative Co Evolutionary Algorithm(CCEA) is designed. The new CCEA can optimize two BRBs and weight coefficient simultaneously, and can improve the diagnostic performance significantly.