Investigation Of Deep Learning-based Joint Segmentation And Registration Algorithm For Infant Brain

In the infancy stage,behavior and cognition grow rapidly,with the fast development of brain morphology.It is critical to accurately segment the infant brain tissues for studying early infant brain developmental patterns,the corresponding neural developmental mechanism of behavior and cognition,and brain diseases.Due to the difficulties such as limited acquisition time,poor imaging quality,high professional requirements for accurate delineation of various brain tissues,manual labeling of large quantities of data requires lots of time and effort.Moreover,there will be inconsistencies in the standards of manual segmentation.Thus higher requirements for infant brain segmentation and registration are needed.In this thesis,systematic research and discussion were conducted on the deep-learning-based segmentation,registration,as well as joint segmentation and registration algorithms for the infant brain.The following were summaries of the conducted research work.(1)Analyze the accuracy and generalizability of DenseNet based segmentation algorithms of infant brain,and determine the best comprehensive performance algorithm.Due to the quick maturation of infant brain tissue,even for MR images within a small age range(e.g.,from newborn to 6 months),there exists huge differences in brain size and image contrast among different subjects and across different tissues.Therefore,it is challenging to develop infant brain segmentation algorithms which are generalizable for images from multi-sites with different ages or/and acquisition protocols.Studies indicated that deep-learning-based approaches performed better in isointensity infant brain(6-9 months old)segmentation,with low image contrast between gray and white matter.However,it has been rarely considered that whether deep-learning-based methods could be generalized to brain image segmentation of infants at other ages,such as neonatal babies.Infant data from iSeg-2017 and UNC neonatal brain images were used to systematically explore the accuracy and generalizability of the deep-learning-based infant brain tissue segmentation algorithms,which are considered as the state-of-the-art methods with better performance compared to non-learning-based conventional methods.The testing results were compared to evaluate the segmentation accuracy of the algorithms when trained and tested within the same cohort;meanwhile,the testing results were quantified and evaluated for the generalizability when trained and tested using different cohorts.Experimental results showed that HyperDenseNet outperformed among all evaluated methods.By taking account of the performance from various aspects,our findings provide conclusive guidance for developing segmentation algorithms with deep-learning models,which are accurate and generalizable for segmenting infant brain images with different acquisition parameters and scanned at different ages.(2)By analyzing the deep-learning-based segmentation and registration algorithms on infant brain,a new deep-learning-based joint segmentation-registration algorithm was designed.It is known that image segmentation and registration were highly associated.In the joint segmentation-registration algorithm,segmentation result warped from labeled image to unlabeled image by registration model provided topological structure information for unlabeled image,which could help to improve the segmentation performance.Meanwhile,the more accurate on segmentation,the better results on registration.Hence,segmentation and registration models could adequately take advantage of,and benefit from each other.However,few studies were performed to jointly use deep-learning-based image registration and segmentation.So a new deeplearning-based joint segmentation-registration algorithm trained with partially labeled data was proposed,in which the predicted segmentation result could be produced from the segmentation and registration model,served as a prior information to guide the model training.VoxelMorph was known as one of the most popular deep-learningbased methods for brain registration.Considering that the weakly-supervised VoxelMorph model was originally used for adult brain registration,experiments were first carried out to select the hyperparameter for UNC infant data.Then,weaklysupervised VoxelMorph model was combined with HyperDenseNet model,and experiments on UNC neonatal brain data were performed.Experimental results revealed that the dice similarity coefficient of the warped segmentation in the joint segmentation-registration model was 1.5%higher,comparing with the separate registration model;and the topological structure of segmentation results was more accurate than the separate segmentation model.This work has proved that the joint segmentation-registration algorithm generate better performance on the infant brain segmentation and registration.In the future,the framework of this algorithm could be extended to other applications such as the segmentation and reconstruction of fetal brain images,in which segmentation and registration are more difficult,whereas high accuracy is required,yet the manual-labelling ground truth is limited.