8:30am - 8:45am
A 3D Microscopic Journey to the Lung: One of the Most Complexly Structured Mammalian Organs
1Institute of Anatomy, University of Bern, Switzerland; 2Swiss Light Source, Paul Scherrer Institute, Villigen, Switzerland; 3Institute for Biomedical Engineering, University of Zurich and ETH Zurich, Switzerland
The respiratory zone of the mammalian lung is divided into functional units called acini. A detailed 3D model is needed in order to complete our understanding on acinar microscopic development and dynamics during breathing. In the past, many attempts have been done using 2D discontiguous methods but only recent advances in synchrotron based X-ray tomographic microscopy (SRXTM) allowed us to reach the required micrometer range resolution in 3D. However the price we pay are increased noise, low contrast and a limited field of view. Thus further steps must be taken before drawing biologically relevant conclusions.
Utilizing phase contrast SRXTM at the TOMCAT beamline we imaged fresh post mortem adult mice at different lung internal pressure level to study acinar dynamics. In addition we imaged extracted lung lobes of rats at different post-natal time-points using absorption SRXTM to examine lung development. For our post mortem scans we demonstrate how to increase their contrast. We propose an automatized graph cut based approach that increases their field of view and mitigates image-blending artifacts. An automated 3D segmentation algorithm that is able to differentiate unambiguously between air and lung parenchyma is presented. Finally we propose a fractal dimension descriptor to quantify the acinar structural complexity and quantify differences between days of development.
Our analysis shows statistically significant differing 3D heterogeneous distension patterns during ventilation. We postulate a complex deformation that is coupled to the pressure level and the acinar location within the thorax. Our extracted fractal dimensions indicate that the shape of acini varies for a given age. An additional difference between post-natal day 4 and post-natal day 60 indicates that complexity of the acinar airways increases during development. We hypothesize that non-uniform branching occurs or new structures are formed and consequently changing the fractal dimension.
8:45am - 9:00am
Natural Gradient for Spiking Neurons
Department of Physiology, University of Bern, Switzerland
While in reinforcement learning, optimal synaptic plasticity follows the reward gradient (Williams, Mach. Learn., 1992), in supervised learning, it follows the negative error gradient. Various parameters can be adapted within a synapse such as the amplitude of the postsynaptic potential at the sub-synaptic site, the postsynaptic receptor density, or e.g. the number of presynaptic release sites. While changing these parameters in the direction of their local gradient, the synaptic contribution to the somatic potential also changes. These somatic potential contributions eventually determine the ring behavior, and one may directly derive the gradient rule for these postsynaptic potentials (PSP) produced by the individual synapses.
For the theory to be consistent, it should not matter whether the optimal PSP changes originate in an optimal change of the number of presynaptic release sites, the postsynaptic receptor densities, or the local potential amplitudes. Yet, the local gradients induce dierent PSP changes, and hence the theory lacks consistency. To resolve the issue, the standard «Euclidean» gradient calculations need to be replaced by the natural gradient (Amari, Neural Comput., 1998).
We derive the natural gradient learning rule for spiking neurons. Instead of the Euclidean metric, this rule calculates the gradient with respect to a local metric on the synaptic weight space that is given intrinsically by the neuron's input-output distribution. As compared to the Euclidean gradient, the natural gradient shows an input rate dependent scaling of the individual synapses, and an additional hetero-synaptic term. In addition, it is coordinate invariant and hence independent of which specic synaptic parameters are adapted during plasticity.
9:00am - 9:15am
Investigation of Cardiopulmonary Exercise Testing Using a Dynamic Leg Press and Comparison with a Cycle Ergometer
1ARTORG Center for Biomedical Engineering Research, University of Bern, Switzerland; 2Institute for Rehabilitation and Performance Technology, Bern University of Applied Sciences, Switzerland
Purpose: Leg-press machines are widely employed for musculoskeletal conditioning of the lower-limbs and they provide cardiovascular benefits for resistance training in cardiac patients. The aim of this study was to assess the feasibility of a dynamic leg press (DLP) for incremental cardiopulmonary exercise testing (CPET) and to compare the results with those obtained using a cycle ergometer (CE).
Methods: Twelve able-bodied subjects aged 27 ± 4 years (mean ± standard deviation) performed incremental cardiopulmonary exercise tests on a DLP and on a CE. To facilitate CPET, the DLP was augmented with force and angle sensors, a work rate estimation algorithm, and a visual feedback system. Gas exchange variables and heart rate were recorded breath-by-breath using a cardiopulmonary monitoring system.
Results: Peak oxygen uptake and peak heart rate were significantly lower for the DLP than for the CE: peak oxygen uptake was 3.2 ± 0.5 vs. 4.1 ± 0.5 L/min (DLP vs. CE, p = 6.7 x 10-6); peak heart rate was 174 ± 14 vs. 182 ± 13 bpm (DLP vs. CE, p = 0.0016). Likewise, the sub-maximal cardiopulmonary parameters, viz. the first and second ventilatory thresholds, and ramp duration were significantly lower for the DLP.
Conclusions: The dynamic leg press was found to be feasible for CPET: the approach was technically implementable and all peak and sub-maximal cardiopulmonary parameters were able to be identified. The lower outcome values observed with the DLP can be attributed to a peripheral factor, namely the earlier onset of muscular fatigue.
Keywords: cardiopulmonary exercise testing; dynamic leg press; cycle ergometer; oxygen uptake; heart rate; ventilatory threshold.
9:15am - 9:30am
Segmentation of Pathological Lung Tissue Using Deep Convolutional Neural Networks
1ARTORG Center for Biomedical Engineering Research, University of Bern, Switzerland; 2Department of Diagnostic, Interventional and Pediatric Radiology, Inselspital, Bern University Hospital, Switzerland; 3Department of Emergency Medicine, Inselspital, Bern University Hospital, Switzerland
The detection of pathologies in HRCT scans with interstitial lung disease (ILD), is a basic component towards the automatic quantification of ILDs and a valuable identifier for the final diagnosis. Scope of this study is the development and evaluation of an automatic system for the recognition and segmentation of pathological ILD lung tissue. The system is based on deep learning techniques, which have recently achieved impressive results in a variety of problems. Specifically, we propose and evaluate a deep convolutional neural network (CNN), designed for the segmentation of ILD patterns.
Based on two multimedia ILD databases (University Hospital of Geneva, University Hospital of Bern) consisting of 172 unique HRCT scans, and a total area of 211.4*105 annotated pixels, 2575 annotated slices were extracted. Six lung patterns were considered: normal, ground glass opacity (GGO), micronodules, consolidation, reticulation and honeycombing. To this end, a CNN that was designed particularly for ILD pattern segmentation was trained in an end-to-end and semi-supervised manner. For the evaluation of the results a cross-validation scheme was adopted, where the cases were split on a patient level. The system reached a performance of nearly 82% in terms of accuracy, demonstrating the potential of CNNs in analyzing lung pathological tissue.
The CNN showed very promising results in lung pattern recognition, outperforming state-of-the-art methods. Future work includes the integration of this system into a pipeline that will provide a differential diagnosis for ILDs.