The Research On Tumor Co-segmentation Using Multi-modal Images

Cancer is a kind of common malignant and serious disease. Image guided radiotherapy is generally adopted for cancer patients without widely tumor metastasis. In this treatment, the cross-sectional images of patients are needed to locate the boundary of tumors and the organs at risk. With that, dosage distribution of the patient’s body are calculated, and radiotherapy plan are formulated.At present, the radiotherapy targets are often sketched manually in hospital when formulating a radiotherapy plan. Doctors have to check patients’cross-sectional images layer by layer and contour the boundary of the tumors in sequences of images by professional radiotherapy planning software. During the process of sketch, doctors should not only delineate the target accurately but also avoid normal tissues and vital organs are include in the target zone. However, this kind of process has many disadvantages:first, hundreds of cross-sectional images and the possibilities of tumor metastasis in several parts of human body directly indicate that, this is a time consuming and boring task. Then, some kinds of tumors don’t show apparent boundries in images(for example, Nasopharyngeal tumor in CT images). Under these circumstances, anatomy changes caused by tumor growth are considered, and provide and important clue to approximate tumor boundry during doctors’ contouring. Even doctors wiht good knowledge of anatomy have to think over and over again before reach a consensus with an abundance of caution. Therefore, the sketching process by hand is boring, difficult and time-consuming, and the contouring results are very subjective and low reproducible. The study of the methods for automatical or semi-automatical segmentation of tumor has became an important research topic in the field of the image-guided radiation therapy technology, and offered a solution to help doctors to delineat the tumor targets with the higher speed and repeatability.There are many traditional methods to achive medical image segmentation, including simple methods, like threshold and region-grow, and complicate methods, like ASM (Active shape modle) and Graph Cut. However, these classical methods usually use single modality. In recent years, multi-modal images have been widly used in clinlical diagnoses, and this fact provide solid basis of the application/research and development opportunity for multi-modal image segmentation methods. Because single modal image can provide limited information about desease, traditional methods cannot yield accurate result to satisfy the clinical demands, but multi-modal images can provied more complementary information of the lesion and around tissues. Thus, theoretically, the segmentation results of the target area should be more accurate when using multi-modal images.However, even with multimodal images, image segmentation is quite a challenging task because the pathological process responsible for the creation and growth of brain tumors is inherently unpredictable. Consequently, the geometric properties of the tumor do not conform to a particular shape/size distribution, which makes the tumor boundaries universally and irregularly distorted. Furthermore, the tumor is heterogeneous and the border is difficult to localize. Another aspect that complicates segmentation is that artifacts and noise can be easily interfused into images during data acquisition. These obstacles make it impossible to use any kind of shape prior on these normal structures to aid in tumor segmentation, and make it difficult to design a fully automatic segmentation method. It is actually a trend to adopt more than one of the methods mentioned above to complement the drawbacks in a single one. This paper will take the segmentation of multi-modal MR brain tumor images and bio-modal PET/CT NPC(Nasal Pharyngeal Cancer) images as entry points to discusse the work I have done:1) This paper present a population-and patient-specific information based method for the segmentation of brain tumors in multimodal MR images. A Gabor filter bank is used to capture the texture properties. To enhance the multi-modal information, the image difference between the two modalities are also considered as a sub-feature. With these features, positive and negative samples are sampled in/outside the tumor automatically in the training set, and manually sampled in/outside the tumor. An optimal distance metric is learned, aimed at improving the discrimination in the feature space by employing Closed-Form Metric Learning. Furthermore, a AdaBoost classifier is introduced to estimate the probabilities of voxels belonging to the target and the background in the projected feature space. Based on this idea, a new cost function is constructed and optimized via Graph Cut. The optimum weight of population-and patient-specific information weight are obtained by traverse the weighting factor and compare the final segmentation results. The results of experiments show that use both population-and patient-specific information of multi-modal images can significantly improve the stability and accuracy of the algorithm.2) To the question of NPC (Nasal Pharyngeal Cancer) segmentation using PET/CT images, this paper performed qualitative and quantitative analysis about the NPC enhancement effect of FDG-PET/CT and Choline-PET/CT. The evaluation results show that Choline-PET/CT may be superior to FDG-PET/CT for determing gross tumor volume in patients, especially the ones with locally advanced NPC. Since the preparation of tracer11C-Choline are more complicate and expensive than18F-FDG, FDG-PET/CT are widely used as clinical tumor tracer for NPC diagnosed, and the image data used for study are easier to obtain. The rest of the paper discussed the topic of bi-modality co-segmentation of NPC using FDG-PET/CT images. Contrapose the main problem of segmentation "lekage" existing in FDG-PET/CT segmentation, the current paper proposed a method to use location distribution map to restrain the segmentation result. Experiments of adding the restrain to the segmentation framework have been performed, including SVM and sparse representation as the core of segmentation framework. The encouraging evaluation results obtained. Follow-up work will concentrate on the consummation of the segmentation framework and the evaluation of the segmentation result of different methods.