Deep Learning for Computational Neuroimaging

Abstract

Non-invasive neuroimaging studies aim to increase our understanding of the brain in health and disease. The extraction and quantification of morphometric measures such as volume or cortical thickness from the incoming stream of data requires efficient, reliable, and accurate computational workflows. However, existing neuroimaging pipelines involve computationally intensive optimization steps and, thus, do not scale well to large cohort studies. Efficient deep learning networks have the potential to revolutionize image analysis but are so far limited in their applicability due to the primary focus on 1.0 mm voxel-based segmentation and insufficient validations. This thesis addresses these issues and contributes fast, accurate, and extensively validated open-source deep learning solutions for the automated processing of structural human brain magnetic resonance images. <br /> With FastSurfer, a full deep learning based alternative to well-established, traditional neuroimaging pipelines is introduced. The methodological contributions include FastSurferCNN, a 2.5D convolutional neural network (CNN) for whole brain segmentation into 95 classes in under 1 minute, and RECON-SURF, a surface reconstruction stream including a novel spectral spherical embedding, fast mapping of the volume segmentations to the surfaces, and extraction of pointwise and regional thickness estimates. FastSurfer reduces processing times while outperforming traditional tools in terms of accuracy, reliability, and sensitivity. The thesis also addresses cortical surface segmentation with a novel view-aggregating polar parameterization network called p3CNN. The view aggregation across different pole axis orientations alleviates distortions introduced by the non-isometric mapping and is shown to outperform spherical CNNs. With the voxel-size independent neural network (VINN) the thesis further introduces a tool for native image segmentation across a range of different voxel sizes.<br /> A novel network-integrated resolution-normalization layer uses a priori knowledge about the image resolution to internally transition between scales. The internal interpolation retains important image information and outperforms traditional scaling augmentation. VINN consistently achieves good results on both, low- and high-resolution images and is highly effective in combating methodological biases in multi-resolution settings. Extending the VINN concept by shifting not only scaling but also rigid transformations into the network, finally gives rise to the VINN with internal augmentations (VINNA). At the first layer scale transition, the multi-dimensional feature maps are flexibly rescaled and randomly transformed to diversify the training distribution. The shift from external to internal augmentations translates to improved accuracy measures in the context of newborn brain segmentation. Overall, the thesis contributions enable reliable, scalable big-data analysis with high robustness for a variety of datasets.