ABSTRACT
In recent years, the rapid development of deep learning approaches has paved the way to explore the underlying factors that explain the data. In particular, several methods have been proposed to learn to identify and disentangle these underlying explanatory factors in order to improve the learning process and model generalization. However, extracting this representation with little or no supervision remains a key challenge in machine learning. In this paper, we provide a theoretical outlook on recent advances in the field of unsupervised representation learning with a focus on auto-encoding-based approaches and on the most well-known supervised disentanglement metrics. We cover the current state-of-the-art methods for learning disentangled representation in an unsupervised manner while pointing out the connection between each method and its added value on disentanglement. Further, we discuss how to quantify disentanglement and present an in-depth analysis of associated metrics. We conclude by carrying out a comparative evaluation of these metrics according to three criteria, (i) modularity, (ii) compactness and (iii) informativeness. Finally, we show that only the Mutual Information Gap score (MIG) meets all three criteria.
ABSTRACT
BACKGROUND AND OBJECTIVE: One of the main issues in the analysis of clinical neonatal brain MRI is the low anisotropic resolution of the data. In most MRI analysis pipelines, data are first re-sampled using interpolation or single image super-resolution techniques and then segmented using (semi-)automated approaches. In other words, image reconstruction and segmentation are then performed separately. In this article, we propose a methodology and a software solution for carrying out simultaneously high-resolution reconstruction and segmentation of brain MRI data. METHODS: Our strategy mainly relies on generative adversarial networks. The network architecture is described in detail. We provide information about its implementation, focusing on the most crucial technical points (whereas complementary details are given in a dedicated GitHub repository). We illustrate the behavior of the proposed method for cortex analysis from neonatal MR images. RESULTS: The results of the method, evaluated quantitatively (Dice, peak signal-to-noise ratio, structural similarity, number of connected components) and qualitatively on a research dataset (dHCP) and a clinical one (Epirmex), emphasize the relevance of the approach, and its ability to take advantage of data-augmentation strategies. CONCLUSIONS: Results emphasize the potential of our proposed method/software with respect to practical medical applications. The method is provided as a freely available software tool, which allows one to carry out his/her own experiments, and involve the method for the super-resolution reconstruction and segmentation of arbitrary cerebral structures from any MR image dataset.
Subject(s)
Image Processing, Computer-Assisted , Magnetic Resonance Imaging , Brain/diagnostic imaging , Female , Humans , Infant, Newborn , Male , Neuroimaging , Signal-To-Noise RatioABSTRACT
The purpose of super-resolution approaches is to overcome the hardware limitations and the clinical requirements of imaging procedures by reconstructing high-resolution images from low-resolution acquisitions using post-processing methods. Super-resolution techniques could have strong impacts on structural magnetic resonance imaging when focusing on cortical surface or fine-scale structure analysis for instance. In this paper, we study deep three-dimensional convolutional neural networks for the super-resolution of brain magnetic resonance imaging data. First, our work delves into the relevance of several factors in the performance of the purely convolutional neural network-based techniques for the monomodal super-resolution: optimization methods, weight initialization, network depth, residual learning, filter size in convolution layers, number of the filters, training patch size and number of training subjects. Second, our study also highlights that one single network can efficiently handle multiple arbitrary scaling factors based on a multiscale training approach. Third, we further extend our super-resolution networks to the multimodal super-resolution using intermodality priors. Fourth, we investigate the impact of transfer learning skills onto super-resolution performance in terms of generalization among different datasets. Lastly, the learnt models are used to enhance real clinical low-resolution images. Results tend to demonstrate the potential of deep neural networks with respect to practical medical image applications.