Transfer Learning Based ROI Region Extraction In Stomach CT Image Sequences

With the rapid development of scientific technology, human health problem has gradually become the focus of attention. Gastric cancer is a high incidence tumor, which threats seriously the health of human life. As an important factor for patients’survival rates, lymph nodes often appear near the stomach for about5cm. Only these regions of interesting are accurately segmented from the whole images, the doctor could carry on the subsequent detection effectively and make a right decision for the diagnosis. The main works of this thesis are summarized as follows:(1)Because the existing interactive image segmentation methods are time-consuming and its effects can not reach the requirements of practical application, a new method for watershed region merge based stomach image segmentation is put forward. The proposed method draw lines to mark some object and background regions, and they define inner procuct of gray gradient feature vector as the similarity criterion. Compared with most of the traditionnal methods, the segmentation accuracy of the proposed method is greatly improved.(2)A seed point transfer growth stomach CT image sequences segmentation method is proposed. In order to extract all the object regions from a complete CT sequence, several seeds are selected manually in the object region for the first image, a global scope similarity search according to the selected seeds and a given threshold to get a binary result, the center and its neighbor of the result are transfered to the adjacent image as its seeds.The proposed method not only improves the speed of segmentation but also overcomes the problem of owe segmentation.(3) Because of the weak edge of the object region in stomach CT image, the existing active contour sequence segmentation models often appear edge leak problem. So a contour transfer evolution based stomach CT sequence segmentation method is proposed. This method combines the region based model and edge model based perfectly, which transfers the growth edge of the object region in current image to the adjacent image as its initial contour. As a result, the new model reduces iteration times and improves the accuracy of the segmentation results.