Medical Image Techniques Based On Registration, Fusion And Segmentation And Its Application In Precision Medicine Of Lung Cancer

Precision medicine technology develops rapidly since 21th century and it is officially included in the 13th Five-Year Plan. Meanwhile, the advanced medical equipment plays an important role in precision medicine of cancer. The multimodal medical image data captured by medical equipment can provide clinicians with more abundant and accurate patient information using modern digital image processing technology. Further, it will contribute significantly to making a precise diagnosis and the treatment plan. This thesis studies the medical image registration, fusion and segmentation, and the research results are applied to the precision medicine of lung cancer based on PET and CT images. The main work includes:(1) Proposing a novel medical image registration method based on mixed mutual information and improved particle optimization algorithm. During each iteration of the proposed algorithm, the improved particle swarm optimization algorithm based on Renyi's entropy was adopted firstly in global searching phase. Then the mutual information measure based on Shannon's entropy was taken as the objective function while the Powell algorithm was used to obtain the optimal solution in local searching phase. In the experiment of mono-modality and multi-modality medical image registration, the proposed algorithm has the advantage in accuracy and effectiveness. The algorithm is applied to the registration of lung cancer based on PET and CT images, it has better performance on the evaluation of subjective and objective.(2) Proposing a medical image fusion algorithm based on projective dictionary pair learning. For training samples acquired by preprocessing technology, the algorithm will obtain jointly a synthesis dictionary and an analysis dictionary by projective dictionary pair learning and sparse coding of training samples using simultaneous orthogonal matching persuit. The sparse coding fusion of high frequency using the summation method and the low frequency is reconstructed by the weighted average method. The fusion results based on CT and MRI images of brain show that the proposed algorithm can effectively preserve the anatomical features and enhance greatly the image sharpness and contrast. According to the judgement of the expert, the fused image of registered PET and CT images of lung cancer effectively retain the feature of the source images and easily distinguish the pathological field from the normal filed. The objective evaluation further validate the assessment of the clinician.(3) Proposing a medical segmentation algorithm based on joint fitting energy model with local structure tensor. The algorithm utilizes the global fitting energy with spatially varying mean and variance, which is beneficial and significant to reduce the sensitivity to the initialization of the contour and speed up the convergence rate, and incorporate the local region fitting energy described by the structure tensor, which is effective for the image with intensity inhomogeneity and preserve the structure of the boundaries. Compared with the state-of-the-art segmentation methods, the experimental results demonstrate the effectiveness and robustness of the proposed algorithm. The proposed algorithm is proved by the segmentation of fused image of lung cancer to have a high accuracy and its vital value in the precise diagnosis and the target delineation with lung cancer.