Lorenz Berger

Lorenz Berger, Eoin Hyde, Matt Gibb et al.

Training deep neural networks on large and sparse datasets is still challenging and can require large amounts of computation and memory. In this work, we address the task of performing semantic segmentation on large volumetric data sets, such as CT scans. Our contribution is threefold: 1) We propose a boosted sampling scheme that uses a-posterior error maps, generated throughout training, to focus sampling on difficult regions, resulting in a more informative loss. This results in a significant training speed up and improves learning performance for image segmentation. 2) We propose a novel algorithm for boosting the SGD learning rate schedule by adaptively increasing and lowering the learning rate, avoiding the need for extensive hyperparameter tuning. 3) We show that our method is able to attain new state-of-the-art results on the VISCERAL Anatomy benchmark.

Luis C. Garcia-Peraza-Herrera, Martin Everson, Wenqi Li et al.

In this work, we have concentrated our efforts on the interpretability of classification results coming from a fully convolutional neural network. Motivated by the classification of oesophageal tissue for real-time detection of early squamous neoplasia, the most frequent kind of oesophageal cancer in Asia, we present a new dataset and a novel deep learning method that by means of deep supervision and a newly introduced concept, the embedded Class Activation Map (eCAM), focuses on the interpretability of results as a design constraint of a convolutional network. We present a new approach to visualise attention that aims to give some insights on those areas of the oesophageal tissue that lead a network to conclude that the images belong to a particular class and compare them with those visual features employed by clinicians to produce a clinical diagnosis. In comparison to a baseline method which does not feature deep supervision but provides attention by grafting Class Activation Maps, we improve the F1-score from 87.3% to 92.7% and provide more detailed attention maps.

Lorenz Berger, Eoin Hyde, M. Jorge Cardoso et al.

Deep convolutional neural networks (CNNs) have shown excellent performance in object recognition tasks and dense classification problems such as semantic segmentation. However, training deep neural networks on large and sparse datasets is still challenging and can require large amounts of computation and memory. In this work, we address the task of performing semantic segmentation on large data sets, such as three-dimensional medical images. We propose an adaptive sampling scheme that uses a-posterior error maps, generated throughout training, to focus sampling on difficult regions, resulting in improved learning. Our contribution is threefold: 1) We give a detailed description of the proposed sampling algorithm to speed up and improve learning performance on large images. We propose a deep dual path CNN that captures information at fine and coarse scales, resulting in a network with a large field of view and high resolution outputs. We show that our method is able to attain new state-of-the-art results on the VISCERAL Anatomy benchmark.

Lorenz Berger, Rafel Bordas, David Kay et al.

A stabilized conforming mixed finite element method for the three-field (displacement, fluid flux and pressure) poroelasticity problem is developed and analyzed. We use the lowest possible approximation order, namely piecewise constant approximation for the pressure and piecewise linear continuous elements for the displacements and fluid flux. By applying a local pressure jump stabilization term to the mass conservation equation we ensure stability and avoid pressure oscillations. Importantly, the discretization leads to a symmetric linear system. For the fully discretized problem we prove existence and uniqueness, an energy estimate and an optimal a-priori error estimate, including an error estimate for the divergence of the fluid flux. Numerical experiments in 2D and 3D illustrate the convergence of the method, show the effectiveness of the method to overcome spurious pressure oscillations, and evaluate the added mass effect of the stabilization term.

Lorenz Berger

In this thesis we develop a stabilised finite element method for solving the equations of poroelasticity to enable solving complex models of biological tissues such as the human lungs. For the proposed numerical scheme, we use the lowest possible approximation order: piecewise constant approximation for the pressure, and piecewise linear continuous elements for the displacements and fluid flux. Due to the discontinuous pressure approximation, sharp pressure gradients due to changes in material coefficients or boundary layer solutions can be captured reliably. We begin by developing theoretical results for approximating the linear poroelastic equations valid in small deformations. In particular, we prove existence and uniqueness, an energy estimate and an optimal a-priori error estimate for the discretised problem. We then extend this work and construct a stabilised finite element method to solve the poroelastic equations valid in large deformations. We present the linearisation and discretisation for this nonlinear problem, and give a detailed account of the implementation. We rigorously test both the linear and nonlinear finite element method using numerous test problems to verify theoretical stability and convergence results, and the method's ability to reliably capture steep pressure gradients. Finally, we derive a poroelastic model for lung parenchyma coupled to an airway fluid network model, and develop a stable method to solve the coupled model. Numerical simulations, on a realistic lung geometry, illustrate the coupling between the poroelastic medium and the network flow model, and simulations of tidal breathing are shown to reproduce global physiologically realistic measurements. We also investigate the effect of airway constriction and tissue weakening on the ventilation, tissue stress and alveolar pressure distribution.