Automatic PET/CT Image Segmentation Based On Multi-atlas Method

With the advantages of prominent imaging techniques in both functional metabolic imaging and anatomical imaging,Positron Emission Tomography/Computed Tomography(PET/CT)imaging is widely used in diagnosis or treatment of cancer.In the application process of human torso PET/CT images,the segmentation of multiple organs of the human body is the prerequisite step of clinical diagnosis or treatment.Therefore,the accuracy and efficiency of organ segmentation is crucial for the success of subsequent image analysis.PET/CT images are characterized by poor image contrast and blurred organ edges.The traditional manual segmentation requires tedious human intervention and long-time professional training,and the segmentation quality depend on the experience and working status of the workers.Therefore,we introduce the prior information of anatomy through multi-atlas algorithm to achieve the purpose of accurate and efficient multi-organ segmentation of human torso PET/CT image.We firstly studied the atlas selection strategies based on low-dose CT images.As a key prerequisite step of multi-atlas segmentation,atlas selection determines the accuracy of subsequent organ segmentation.When the selected atlases are close enough to the samples,near optimal segmentation result can be obtained.We evaluated types of selection criteria based on image similarity and physiological standards.The experimental results show that the Body Mass Index strategy is slightly inferior to the image similarity strategy in terms of accuracy,but it takes less computation time and is more suitable for large segmentation work.Segmentation results show that some organs,like skeleton,lungs,heart,and liver,do not require atlas strategy studies.The accuracy of spleen and kidneys were imperfect for all selection strategies and should be focused on in subsequent studies.Based on the evaluation of atlas selection strategies,we propose to improve the segmentation accuracy of abdominal organs and the segmentation efficiency based on dual-modality PET/CT image.Due to the poor image contrast,the abdominal organs in CT are difficult to segment.We use the PET and CT images to perform segmentation based on the high contrast of abdominal organs in PET images.The algorithm uses MAS workflows based on CT and PET images,respectively.We use dual-resolution coarse-to-fine procedure to segment multiple organs.This study also replaces the traditional voxel label with curved multi-atlas segmentation to improve the algorithm speed and edge smoothness.Experimental results show that the use of dual-modality images obtains significantly more accurate results than single-modal segmentation(median Dice similarity coefficient over 0.8 and median Average surface distance less than 2.5mm).By replacing the traditional voxel label atlas with point cloud-based surface multi-atlas segmentation,the algorithm significantly improved computation time(about 10 minutes for each target)and edge smoothness of the segmented organ.In addition,we segmented a large number of PET/CT images using the proposed algorithm,and constructed an online anatomy medical education system which provides inter-subject anatomical differences and morphological changes.