Swiss Medical Image Computing Day 2015

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

When and Where

The workshop will take place on November 24th, 2015 at the ARTORG Center (Room F502), University of Bern (Murtenstrasse 50, 3008, Bern).

Registration

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

Schedule

09:45 - 10:00 Opening remarks  
10:00 - 10:30
Vimal Chandran
(ISTB, Uni. Bern)
Prediction of Trabecular Bone Anisotropy from Quantitative Computed Tomography using Supervised Learning and a Novel Morphometric Feature Descriptor
10:30 - 11:00
David Romascano
(EPFL)
Advanced tissue microstructure characterization using diffusion MRI data
11:00 - 11:30
Coffee break
 
11:30 - 12:30
Keynote Talk
Dr. Danail Stoyanov
(UCL, UK)
Inferring Geometry+ from Endoscopic Video
12:30 - 13:30
Lunch break
 
13:30 - 14:00
Ksenia Konyushkova
(EPFL)
Introducing Geometry into Active Learning for Image Segmentation
14:00 - 14:30
Natalia Chicherova
(Uni. Basel)
Efficient Algorithm for elastic 2D-Histology to 3D-Data Registration
14:30 - 15:00
Stergios Christodoulidis
(ARTORG, Uni. Bern)
Lung pattern analysis using machine learning for the diagnosis support of interstitial lung diseases
15:00 - 15:30
Coffee break
 
15:30 - 16:00
Thomas Gerig
(Uni. Basel)

Gaussian Process Morphable Models

16:00 - 16:30

Mario Fartaria De Oliveira 
(Uni. Lausanne)

Automated detection of white matter and cortical lesions in early-stages of multiple sclerosis

16:30 - 17:00
Dimitrios Damopoulos
(ISTB, Uni. Bern)

Automatic Detection of Lumbar Vertebrae from CT Images Using an AdaBoost Classifier and Haar-Like Features


Abstracts

 

Inferring Geometry+ from Endoscopic Video

(Dr. Danail Stoyanov, UCL, UK)
The growing uptake of minimally invasive surgery has opened exciting opportunities for the application of computer vision techniques on images taken by cameras inserted into the patient’s anatomy. In this talk, I will discuss some of our past and ongoing work in inferring 3D geometry and motion from monocular and stereo endoscopic video. Computer assisted interventions can benefit from such methods by enhancing visualisation at the surgical site, as well as, the surgeon’s navigation within the anatomy and potentially surgical dexterity when coupled with robotic instrumentation. We have investigated various possible applications for vision approaches in robotic assisted surgery and computer assisted interventions and I will discuss some of the potential clinical needs and our work in different surgical specialisations.

 

Prediction of Trabecular Bone Anisotropy from Quantitative Computed Tomography using Supervised Learning and a Novel Morphometric Feature Descriptor 

(Vimal Chandran - ISTB, Uni. Bern)
Patient-specific biomechanical models including local bone mineral density and anisotropy have gained importance for assessing musculoskeletal disorders. However the trabecular bone anisotropy cap- tured by high-resolution imaging is only available at the peripheral skele- ton in clinical practice. In this work, we propose a supervised learning approach to predict trabecular bone anisotropy that builds on a novel set of pose invariant feature descriptors. The statistical relationship between trabecular bone anisotropy and feature descriptors were learned from a database of pairs of high resolution QCT and clinical QCT reconstruc- tions. On a set of leave-one-out experiments, we compared the accuracy of the proposed approach to previous ones, and report a mean predic- tion error of 6% for the tensor norm, 6% for the degree of anisotropy and 19◦ for the principal tensor direction. These findings show the po- tential of the proposed approach to predict trabecular bone anisotropy from clinically available QCT images.

Advanced tissue microstructure characterization using diffusion MRI data 

(David Romascano - EPFL) 

Magnetic Resonance Imaging can be used to probe water diffusion in the brain non-invasively. As water diffusion is modulated by its environment, diffusion MRI can be used to infer tissue microstructural properties, like axon trajectories, axonal diameter, fibre dispersion or intra- and extra-cellular volume fractions. We will present our work on providing two flexible frameworks that reformulate diffusion MRI problems into equations that are solvable using convex-optimization algorithms, and how they can be used to fit advanced white matter models.

Automatic Detection of Lumbar Vertebrae from CT Images Using an AdaBoost Classifier and Haar-Like Features

(Dimitrios Damopoulos ISTB, Uni. Bern)
In this talk, I will present my progress so far on my PhD project, which ultimately aims for the automatic segmentation of vertebrae in medical images and the identification of present pathologies. More specifically, I will present a pipeline for the identification of the centers of vertebrae in CT images of the lumbar spine. The pipeline exploits the classical Viola-Jones detection algorithm, adapted for the 3-D space of the CT images and followed by a robust post-processing step. More specifically, we first parse the sagittal planes of the CT scan using an AdaBoost classifier trained on Haar-like features. We then apply the mean shift clustering technique in order to find the peaks of the classification score of the AdaBoost classifier. Every such peak represents a candidate position for the center of a vertebra. We then proceed by extracting the coronal planes around the candidate positions. We eliminate any possible false positive detection in the sagittal plane by parsing the coronal views of the candidates with a second AdaBoost classifier, trained to detect the centers of the coronal view of the vertebrae. At the end of the pipeline, we have the coordinates of the centers of vertebra in the anatomical axis and a rectangular cuboid that bounds the detected vertebra region.

Automated detection of white matter and cortical lesions in early-stages of multiple sclerosis 

(Mario Fartaria De Oliveira - Uni. Lausanne)
Early detection of Multiple Sclerosis (MS) lesions is critical for patients diagnosis, follow-up and optimal treatment. Today, advanced Magnetic Resonance Imaging (MRI) sequences provide improved spatial resolution and intensity contrast for MS lesions visualization, both in the cortex and subcortical white matter (WM). In the last years, several groups have proposed image processing methods to detect and segment MS lesions based on clinical conventional MRI sequences. However, existing automated methods are dedicated to WM lesions and applied only to MS patients with long disease duration. We developed a lesion detection method, based on the k-nearest neighbour (k-NN) technique, to detect lesions as small as 0.0036 mL. This method uses the image intensity information from up to four 3D MRI sequences. Two of them are conventionally used in clinical routines (magnetization-prepared rapid gradient-echo, MPRAGE; and 3D fluid-attenuated inversion recovery, FLAIR) and the others are non-conventional sequences used in research fields (magnetization-prepared two inversion-contrast rapid gradient-echo, MP2RAGE; and 3D double-inversion recovery, DIR). To these intensity features we added the information obtained by the spatial coordinates from Montreal Neurological Institute (MNI) framework and tissue prior probabilities provided by the International Consortium for Brain Mapping (ICBM). Quantitative assessment was done in large data set of 39 early-stages MS patients with a leave-one-out cross-validation where ground truth lesion map was based on manual segmentation of lesions performed by two experts. Results will show that the use of non-conventional sequences improve the detection of small lesions, particularly cortical lesions.

Introducing Geometry into Active Learning for Image Segmentation 

(Ksenia Konyushkova - EPFL)
We propose an Active Learning approach to training a segmentation classifier that exploits geometric priors to streamline the annotation process in 3D image volumes. To this end, we use these priors not only to select voxels most in need of annotation but to guarantee that they lie on 2D planar patch, which makes it much easier to annotate than if they were randomly distributed in the volume. A simplified version of this approach is effective in natural 2D images. We evaluated our approach on Electron Microscopy and Magnetic Resonance image volumes, as well as on natural images. Comparing our approach against several accepted baselines demonstrates a marked performance increase.
 

Efficient Algorithm for elastic 2D-Histology to 3D-Data Registration 

(Natalia Chicherova - Uni. Basel)
Localizing a histological slice in a three-dimensional dataset is a challenging 2D-3D registration problem. Recently we developed an automatic algorithm that could successfully find the best matching position of histological section in micro computed tomography (μCT) data. For the majority of the datasets, the result of localizing was excellent. However, for some datasets the best matching section was slightly off the ground-truth position. Here, we introduce an additional optimization framework based on mutual information which enabled us accurately register histology to the volume data. The average improvements of the plane coordinates for ten data sets were: for the polar angle 0.33°, azimuthal 22° and slice position along the specimen 10 slices.

 

 

Lung pattern analysis using machine learning for the diagnosis support of interstitial lung diseases 

(Stergios Christodoulidis - ARTORG, Uni. Bern)
Interstitial lung diseases (ILDs) is a group of more than 200 chronic lung disorders characterized by scarring (i.e. fibrosis) and/or inflammation of the interstitium that cause respiratory failure. It is often the case for ILDs that additional invasive procedures should be considered for the final diagnosis such as bronchoalveolar lavage or histological confirmation.That is due to the large quantity of radiological data that radiologists have to scrutinize and the resemblance between the different ILD findings. Thus, in order to avoid the dangerous histological biopsies, the interpretation of HRCT imaging data in an automatic manner is a really crucial task towards a computer aided (CAD) system for the ILD diagnosis. The aim of this project is to design, develop and evaluate a computer aided diagnosis (CAD) system that will assist clinicians with the diagnosis of ILDs without the need of surgical biopsies. This talk briefly presents an outline of the proposed CAD system while focuses mostly on the lung pattern classification component. More specifically, a comparative assessment of different schemes for lung pattern classification is presented where these schemes range from the combination of classical texture descriptors and classifiers till more recent systems that utilize unsupervised learning techniques.

 

Gaussian Process Morphable Models 

(Thomas Gerig - Uni. Basel)
Models of shape variations have become a central component for the automated analysis of images. Statistical shape models (SSMs) represent a class of shapes as a normal distribution, whose parameters are estimated from example shapes. We propose a generalization of SSMs, which we refer to as Gaussian Process Morphable Models (GPMMs). We model the shape variations with a Gaussian process and a covariance function. Possible models include commonly used priors for image registration, such as the B-spline model, or models learned from data, such as the standard SSM. Using medical applications, such as registration and segmentation, we show the flexibility of combining different models using Gaussian processes.

 

Contact Raphael Sznitman for any questions.

This scientific meeting is in part funded by the University of Bern, Fund for the Promotion of Young Researches.