Auxiliary Diagnosis And Treatment For Lung Tumor Based On PET/CT Image Processing

Lung tumor as the most deadly malignant tumor in the world, has brought tremendous threats to the health of mankind. PET/CT technology plays an important role in diagnosis and treatment of lung tumor. This thesis focuses on the auxiliary diagnosis and treatment of lung tumor based on PET/CT image processing, including the image fusion, the image segmentation and the tumor growth model. These technologies allows medical doctors to efficiently diagnose, delineate tumor and revise the radiotherapy schema based on the information provided by PET/CT images. The main work of this thesis is as follows:First of all, in order to improve the reliability of tumor delineation and tumor stage, This thesis aimes to break through the bottlenecks of the image fusion on PET image, which has low resolution and blur boundaries. Considering the clinical requests, this thesis proposes a novel fusion method in terms of weight computation. By abstracting the saliency map of the output of the cross bilateral filter for PET image, we obtain the image which reflects the feature of PET image and tumor's boundaries of CT image at the same time. This saliency map is obtained as the weight of PET. The sum of the weight of PET and CT equals to 1. Thus we get the fusion image. Furthermore, we add geodesic distance to construct the kernel of filter to obtain more precise boundary information of CT image, meanwhile the robustness of the parameter choosing is improved. Our algorithm remains both the feature of CT and PET images, and is proved to be effective compared with three other fusion algorithms.Secondly, auto-segmentation methods are applied to evaluate the treatment of radiotherapy, predict the performance of the tumor growth and ensure the consistency of the segmentation. The method releases the medical doctor from heavy workload and minimizes the differences of manual tumor segmentation. The thesis proposes a segmentation algorithm based on the active contour model. The energy function which combines regional-scalable fitting term and local Hausdorff distance more accurately partitions image into target region and background. This algorithm has relatively fast convergence rate and less time consumption, which shows good performance on segmentation of PET and ultrasonic images. Step further, we take the intensity distribution into account, adapt the log-logistic distribution to fit the background distribution, and then apply the distribution to the active contour model. Compared with random walk methods and threshold methods, our algorithm has good advantage for PET image segmentation. Our segmentation result can be used in dose-volume model to constrain the performance of tumor growth prediction.Finally, this thesis studies the tumor evolution model to predict the tumor response on radiotherapy based on PET images. According to the environment influence on tumor cell growth, and the curve feature of the Gompertzian function, we modify the proliferation term in the model which reflects the influence of spatial environment on the cell motion. The model we proposed could timely predict the spatial location of the tumor. Therefore, medical doctors can take advantage of the model to modify the radiotherapy schema and forecast the motion of the tumor cell.