Superpixel Based Multi-Organ Segmentation Of Abdominal Images

As the significant reference for clinicians to diagnose abdominal disease,the effective segmentation of organs in abdominal images is the basis and key step of the focus extraction,three-dimensional reconstruction of abdominal organ and the guidance of abdominal surgery in the later time.Most of the existing methods on abdominal organ segmentation are aimed at a single specific organ,in addition,the fuzzy boundary and high gray similarity between adjacent organs in the abdominal images result in difficulties for the traditional multi-organ segmentation methods on pixel level.While,pre-segmentation results with strong consistency and abundant semantic information can be attained on the basis of superpixel,which can provide fresh ideas for multi-organ segmentation of abdominal images.In this paper,the following research work is around multi-organ segmentation of abdominal image:(1)Aiming at the lack of high-quality and annotated abdominal CT image data sets,a complete multi-organ segmentation data set of abdominal CT images was collected and established,including 20 sets of human abdominal CT images(4103 totally),and the main abdominal organs in the images were labeled with organ regions and boundaries manually.(2)Based on the characteristics of wide distribution of gray scale of medical images and strong correlation between upper and lower layers,a new superpixel segmentation method for medical images was proposed.Through the establishing the mapping relationship between the upper and lower layers of the CT images in the clustering process and improving of the distance measurement method on pixels,the superpixel segmentation method did well in medical images.(3)Extracting various features of superpixels,building random characteristics subspaces and training multiple integrated classifiers,the idea of supervised classification was used to solve the problem of abdominal multi-organ segmentation through transforming semantic segmentation of abdominal organs into a multi-classification problem of superpixel blocks.This method improved the accuracy of abdominal multi-organ semantic segmentation.4)In order to reduce the misclassification when using the low-level features purely,combining with prior theory,Gaussian Prior models for each organ about gray and position were established respectively.The model constrained the classification results of the multiclassification probability extreme learning machine in the process of superpixel block classification.Experimental results showed that the method improved the accuracy and the precision of multi-organ segmentation.