Image-based Feature and Prior Learning with Applications to Volumetric Segmentation of Fat-Water Magnetic Resonance Images

Present work enables an automated volumetric segmentation of multiple tissues on fat-water MR images. These segmentations enable an automated volumetry, morphometry, and quantitative analysis of tissues on the images of a large cohort or longitudinal studies on automated diagnosis or efficient therapy planning. This is achieved by proposing and evaluating multiple segmentation approaches to allow the user to select the best approach according to the application, the desired accuracy and frugality, the explainability, and the transparency. In one approach, a hierarchical quadratic random forest classifier is followed by a stack of multiresolution neighborhood graphs and the graph of a hierarchical conditional random field to form a frugal, explainable, and transparent pipeline with the flexibility and controllability of handcrafted features and the ability to tackle the class imbalance and the overfitting. In another approach, a novel algorithm and a novel objective function are proposed to tackle label uncertainties, class imbalance, and overfitting in optimization of discriminative neural network classifiers of any architecture. The evaluations show the superiority of the proposed objective function to commonly used objective functions and the comparability of the proposed pipeline with the best optimized neural networks despite of its lower computational and data demand.