2nd Swiss Medical Image Computing Day 2016

The Swiss Medical Image Computing Day brings together Swiss researchers working on topics related to medical image computing.

When and Where

Change notice: The workshop will take place on November 24th, 2016 in the Maurice E. Müller-Haus building (Murtenstrasse 35, Bern), Room H810. It is across the street from the ARTORG Center (previous venue). 



The workshop is free but registration to the event is mandatory. Please register here.


09:25 - 09:30 Opening remarks  
09:30 - 09:55
Adrien Besson (EPFL)
Compressed Ultrasound beamforming
09:55 - 10:20
Elena Najdenovska (UNIL)
Robust Thalamic Nuclei Segmentation Method Based on Local Diffusion Magnetic Resonance Properties
10:20 - 10:45
Nico Gorbach (ETHZ)
Pipeline Validation for Connectivity-based Cortex Parcellation
10:45 - 11:15
Coffee break
11:15 - 11:40
Chengwen Chu (UNIBE) Fully automatic localization and segmentation of 3D vertebral bodies from CT/MR images via a learning-based method
11:40 - 12:40
Keynote: Prof. Daniel Rueckert (ICL, UK)
Learning clinically useful information from medical images
12:40 - 14:00
Lunch break
14:00 - 14:25
Stergios Christodoulidis (UNIBE)

Texture Classification Using Deep Convolutional Neural Networks

14:25 - 14:50
Agata Mosinska (EPFL) Interactive learning for delineation of curviliniear structures
14:50 - 15:15
Carlos Shokiche (UNIBE) High-throughput Glomeruli Analysis of µCT Kidney Images Using Tree Priors, Scalable Sparse Computation and crowdsourcing

15:15 - 15:45

Coffee break


15:45 - 16:10

Thomas Bolton (EPFL / UNIGE)

Coupled hidden Markov models for the estimation of resting-state networks activity time courses

16:10 - 16:35

Firat Ozdemir (ETHZ)

Automatic Bone Surface Segmentation from Ultrasound Images Using Learned Features and Graphical Modeling

16:35 - 17:00
Stefanos Apostolopoulos (UNIBE)

Age-related Macular Degeneration detection using Optical Coherence Tomography and Deep Learning


Learning clinically useful information from medical images

(Prof. Dr. Daniel Ruekert - Imperial College London, UK)
This talk will focus on the convergence of computer vision and machine learning techniques for the discovery and quantification of clinically useful information from medical images. The first part of the talk will describe machine learning techniques that can be used for image reconstruction, e.g. the acceleration of MR imaging. The second part will discuss model-based approaches that employ statistical as well as probabilistic approaches for segmentation. In particular, we will focus on segmentation techniques that combine patch-based approaches such as dictionary learning with sparsity to improve the accuracy and robustness of the segmentation approaches.

Compressed ultrasound beamforming 

(Adrien Besson - EPFL)
Classical ultrasound (US) image reconstruction mainly relies on the well-known Delay-And-Sum (DAS) beamforming for its simplicity and real-time capability. However, DAS requires an extensive number of samples and delay calculations to obtain high-quality images. Compressed ultrasound beamforming (CUB) proposes an alternative to DAS based on the compressed-sensing (CS) framework which aims at reducing the data rate. CS demonstrates that a signal can be perfectly recovered from fewer samples than required by the Nyquist rate if some properties of both the signals under interest and the acquisition system are respected. In order to account for these properties, CUB redesigns both the acquisition and the reconstruction of US images and leads to high quality reconstruction with less than 20% of the data required by DAS. In the talk, some basic principles of CS and US will be introduced. Then, we will describe CUB in light of the CS framework introduced before.  Eventually, benefits of CUB will be demonstrated through simulation and in vivo experiments

Robust Thalamic Nuclei Segmentation Method Based on Local Diffusion Magnetic Resonance Properties

(Elena Najdenovska - UNIL) 

The thalamus is an essential relay in the cortical-subcortical connections. It is characterized by a complex anatomical architecture composed of numerous small nuclei, which mediate the involvement of the thalamus in a wide range of neurological functions. We present a novel framework for segmenting the thalamic nuclei, which explores the orientation distribution functions (ODFs) from diffusion magnetic resonance images at 3 Tesla. The differentiation of the complex intra-thalamic microstructure is improved with the spherical harmonic (SH) representation of the ODFs providing full angular characterization of the diffusion process in each voxel. The clustering was performed using k-means algorithm initialized in a data-driven manner. The method was tested on 35 healthy volunteers and our results showed a robust, reproducible and accurate segmentation of the thalamus in seven nuclei groups. Six of them closely match the anatomy and were labeled as Anterior, Ventral Anterior, Medio-Dorsal, Ventral Latero-Ventral, Ventral Latero-Dorsal and Pulvinar, while the seventh cluster included the Centro-Lateral and the Latero-Posterior nuclei. Results were evaluated both qualitatively, by comparing the segmented nuclei to the histological atlas of Morel, and quantitatively, by measuring the clusters extent and the clusters spatial distribution across subjects and hemispheres. Furthermore, we show the robustness of our approach across different sequences and scanners, as well as intra-subject reproducibility of the segmented clusters, on two scan-rescan additional datasets. We also observed an overlap between the path of the main long-connections tracts passing through the thalamus and the spatial distribution of the nuclei identified with our clustering algorithm. Our approach, based on SH representations of the ODFs outperforms the one based on angular differences between the principle diffusion directions, which is considered so far as state-of-the-art method. Our findings show an anatomically reliable segmentation of the main groups of thalamic nuclei that could be of potential use in many clinical applications.

Pipeline Validation for Connectivity-based Cortex Parcellation

(Nico Gorbach, ETHZ)
White matter connectivity plays a dominant role in brain function and arguably lies at the core of understanding the structure-function relationship in the cerebral cortex. “Connectivity-based cortex parcellation” (CCP), a framework to process structural connectivity information gained from diffusion MRI, identifies cortical subunits that furnish functional inference. This pipeline of algorithms interprets similarity in white matter connectivity as a criterion to group neuronal units. Validation of the CCP- pipeline is critical to gain scientific reliability of the algorithmic processing steps from DTI data to voxel grouping. In this talk I will present a novel information-theoretic principle based upon a trade-off between informativeness and robustness to assess the validity of the CCP pipeline, including fiber tracking and clustering. We ultimately identify a pipeline of algorithms and parameter settings that tolerate more noise and extract more information from the data than their alternatives.

Fully automatic localization and segmentation of 3D vertebral bodies from CT/MR images via a learning-based method

(Chengwen Chu - UNIBE)
In this talk, we present a method to address the problems of fully automatic localization and segmentation of 3D vertebral bodies from CT/MR images. We propose a learning-based, unified random forest regression and classification framework to tackle these two problems. More specifically, in the first stage, the localization of 3D vertebral bodies is solved with random forest regression where we aggregate the votes from a set of randomly sampled image patches to get a probability map of the center of a target vertebral body in a given image. The resultant probability map is then further regularized by Hidden Markov Model (HMM) to eliminate potential ambiguity caused by the neighboring vertebral bodies. The output from the first stage allows us to define a region of interest (ROI) for the segmentation step, where we use random forest classification to estimate the likelihood of a voxel in the ROI being foreground or background. The estimated likelihood is combined with the prior probability, which is learned from a set of training data, to get the posterior probability of the voxel. The segmentation of the target vertebral body is then done by a binary thresholding of the estimated probability. We evaluated the present approach on two openly available datasets: 1) 3D T2-weighted spine MR images from 23 patients and 2) 3D spine CT images from 10 patients. Taking manual segmentation as the ground truth (each MR image contains at least 7 vertebral bodies from T11 to L5 and each CT image contains 5 vertebral bodies from L1 to L5), we evaluated the present approach with leave-one-out experiments. The experimental results demonstrated the efficacy of the present approach.

Interactive Learning for Delineation of Curvilinear structures

(Agata Mosinska - EPFL)
Supervised machine learning algorithms intrinsically require extensive amounts of annotated ground-truth data. To make such methods truly practical, we propose an Active Learning approach suited especially for annotating elongated structures like neurons or blood vessels. It speeds up the process of annotation by up to 80%, thus greatly decreasing the effort. It does so by taking into consideration geometric specificities of the delineation problem and assuming that the structures are usually smooth. Even though machine learning-based methods proved to be very effective, they still make mistakes and as a matter of fact, the current bottleneck in the neuron reconstruction pipeline is error-correction. We will therefore present another interactive algorithm that detects which parts of the reconstruction are not consistent and can possibly contain mistakes. It is then possible to present them to the expert and validate, without requiring him to visual inspect the whole scan. This way, we make the most out of expert knowledge and efficiency of automatic tools.

Coupled hidden Markov models for the estimation of resting-state networks activity time courses

(Thomas Bolton - EPLF / UNIGE)
Spontaneous brain activity in the absence of any task or stimulus can be monitored through the blood oxygenation level dependent (BOLD) functional magnetic resonance imaging (fMRI) signal, which has enabled the discovery of interacting resting-state networks (RSNs) that play major roles in cognition and brain disorders. Recently, connectivity within and across those networks was shown to be non-stationary over time, leading to novel methodological efforts in order to more accurately resolve the brain functional architecture. Surprisingly, however, temporal modeling is so far rarely employed in retrieving the time courses of RSNs activity profiles. Here, we introduce a coupled hidden Markov model (CHMM) framework, in which each RSN is modeled as a latent variable with varying activity degree, and RSNs are enabled to causally influence each other. Transition probabilities are described by a logistic regression framework, in which sparsity of causal influences is achieved through L1 regularization. We first test this CHMM framework on artificially generated causally related time courses, and compare it to other existing modeling approaches. Then, we describe preliminary results obtained on a RS fMRI dataset of healthy subjects, where initial RSN and time course estimates are obtained through state-of-the-art dynamic functional connectivity tools.

Automatic Bone Surface Segmentation from Ultrasound Images Using Learned Features and Graphical Modeling

(Firat Ozdemir - ETHZ)
Bone surface identification and localization in ultrasound have been widely studied in the contexts of computer-assisted orthopedic surgeries, trauma diagnosis, and post-operative follow-up. Nevertheless, the (semi-)automatic bone surface segmentation methods proposed so far either require manual interaction or fail to deliver accuracy fit for clinical purposes. In this presentation, we utilize the physics of ultrasound propagation in human tissue by encoding this in a factor graph formulation for an automatic bone surface segmentation approach. We comparatively evaluate our method on annotated in-vivo ultrasound images of bones from several anatomical locations, including bones in the forearm (radius, ulna), shoulder (acromion, humerus tip), leg (fibula, tibia, malleolus), hip (iliac crest), jaw (mandible, rasmus) and fingers (phalanges). Our method yields a root-mean-square error of 0.59mm, far superior to state-of-the-art approaches.

Texture Classification Using Deep Convolutional Neural Networks

(Stergios Christodoulidis - UNIBE)
Deep learning techniques have recently achieved impressive results in a variety of computer vision tasks, raising expectations that they could be applied in other domains, such as medical image analysis. Many different medical conditions are manifested as texture alternations in a variety of medical imaging modalities. Therefore, a robust texture classifier may be beneficial for many diagnostic systems. This talk briefly presents a texture classification scheme using deep convolutional neural networks (CNN). Specifically, a novel CNN architecture, designed specifically for the problem of texture classification, will be described and a comparative assessment between other architectures and systems will be presented. Classification performance results will be reported on different texture datasets.

Age-related Macular Degeneration detection using Optical Coherence Tomography and Deep Learning

(Stefanos Apostolopoulos - UNIBE)
Age-related Macular Degeneration (AMD) is the leading cause of blindness for people over 50 years of age in Europe, with an estimated 1 out of 9 people to be affected by this disease by the age of 80. Optical Coherence Tomography (OCT) is a non-invasive imaging technique that allows us to acquire micrometer-resolution volumetric scans of the retina. In this talk, we will explore the use of deep learning methods for detecting AMD in OCT scans of early- and late-stage AMD patients and discuss techniques for handling noisy, high-resolution volumetric data.

High-throughput Glomeruli Analysis of µCT Kidney Images Using Tree Priors, Scalable Sparse Computation and crowdsourcing

(Carlos Shokiche - UNIBE)
Kidney-related diseases have incrementally become one major cause of death. Glomeruli are the physiological units in the kidney responsible for the blood filtration. Therefore, their statistics including number and volume, directly describe the efficiency and health state of the kidney. Stereology is the current quantification method relying on histological sectioning, sampling and further 2D analysis, being laborious and sample destructive. New micro-Computed Tomography (mCT) imaging protocols resolute structures down to capillary level. However large-scale glomeruli analysis remains challenging due to object identifiability, allotted memory resources and computational time. We present a methodology for high-throughput glomeruli analysis that incorporates physiological a priori information relating the kidney vasculature with estimates of glomeruli counts. We propose an effective sampling strategy that exploits scalable sparse segmentation of kidney regions for refined estimates of both glomeruli count and volume. We evaluated the proposed approach on a database of mCT datasets yielding a comparable segmentation accuracy as an exhaustive supervised learning method. Furthermore we show the ability of the proposed sampling strategy to result in improved estimates of glomeruli counts and volume without requiring a exhaustive segmentation of the mCT image. We investigate the generation of groundtruth by non-experts based on crowdsourcing platform. The overall approach can potentially be applied to analogous organizations, such as for example the quantification of alveoli in lungs.

Previous events

1st Swiss Medical Image Computing Day



This scientific meeting is in part funded by the University of Bern.

Contact Raphael Sznitman for any questions.