Machine learning for automatic segmentation of neonatal and fetal MR brain images

Abstract

Fetal MRI is widely used to investigate brain development in utero. MR images do not only visualize brain tissue classes and identify abnormalities but are also able to quantify brain development using volumetric measurements and cortical folding. Such analysis allows us to investigate brain development in case of antenatal abnormalities and to compare this quantitatively with the development of healthy infants. Fetal MRI visualizes fetal brain, maternal body and infant’s body. A prerequisite for quantification is the segmentation of the fetal brain into different tissue classes, such as cortex, white matter and gray matter. Performing this segmentation manually is extremely time-consuming and requires a high level of expertise, not only because of the complex convoluted shapes of the different tissues but also owing to the low image resolution and to fetus motion. Moreover, MR intensity values in tissues are not fully homogeneous in some slices and tissue contrasts are difficult to discern. Consequently, automatic brain tissue segmentation is challenging. Neonatal MRI analysis is of great clinical interest to quantitatively measure brain development and to thereby aid in diagnosis and treatment decisions. Besides anomaly and defects in newborns, premature birth is the most frequent complication in neonates with an incident of 1 per 8.6. Premature infants develop the brain ex utero which may lead to brain injury or secondary developmental. Obtaining image segmentations manually in a large-scale study is both subjective and time-consuming. Moreover, neonatal MRI frequently suffers from artifacts caused by the infant’s motion that occurs during scanning. A further challenge in neonatal MRI segmentation is anatomical variation due to differences in brain morphology between 30 and 40 weeks postmenstrual age (PMA) or critical abnormalities that alter brain shape significantly. All these causes of large variations and abnormalities in neonatal brain MR images call for automatic segmentation methods. Recently, machine learning methods and particularly deep learning achieved excellent segmentation performance in neonatal brain segmentation. The major strength of deep learning techniques is their ability to extract features relevant to the task directly from the data. Consequently, there is no need to derive a set of handcrafted features from the image for classification or regression tasks. Instead, a convolutional neural network (CNN) connects the input and output through a number of convolution layers. These convolution layers consist of trainable convolution kernels with weights and biases which are optimized during the training using a loss function and stochastic gradient descent. CNN is the most popular neural network in image processing, classification, and segmentation tasks. Another type of neural network recently widely used in image generation is the generative adversarial network (GAN). GAN composed of two networks, a generator, and a discriminator. These two networks are trained against each other, the generator network trains to fool the discriminator, and the discriminator train to discriminates better between real and fake data made by generator. Therefore, in image generation, the network generates more realistic images.

The objective of this thesis is to develop CNN-based automatic methods to segment neonatal and fetal MR brain images.