J.-Ph. Thiran

Spread spectrum for compressed sensing techniques in magnetic resonance imaging

Magnetic resonance imaging (MRI) probes signals through Fourier measurements. Accelerating the acquisition process is of major interest for various MRI applications. The recent theory of compressed sensing shows that sparse or compressible signals may be reconstructed from a small number of random measurements in a sensing basis incoherent with the sparsity basis. In this context, we advocate the use of a chirp modulation of MRI signals prior to probing an incomplete Fourier coverage, in the perspective of accelerating the acquisition process relative to a standard setting with complete coverage. We analyze the spread spectrum phenomenon related to the modulation and we prove its effectiveness in enhancing the overall quality of image reconstruction. We also study its impact at each scale of decomposition in a wavelet sparsity basis. Our preliminary results rely both on theoretical considerations related to the mutual coherence between the sparsity and sensing bases, as well as on numerical simulations from synthetic signals.

Lower and upper bounds for approximation of the Kullback-Leibler divergence between Gaussian Mixture Models

Many speech technology systems rely on Gaussian Mixture Models (GMMs). The need for a comparison between two GMMs arises in applications such as speaker verification, model selection or parameter estimation. For this purpose, the Kullback-Leibler (KL) divergence is often used. However, since there is no closed form expression to compute it, it can only be approximated. We propose lower and upper bounds for the KL divergence, which lead to a new approximation and interesting insights into previously proposed approximations. An application to the comparison of speaker models also shows how such approximations can be used to validate assumptions on the models.

Modelling human perception of static facial expressions

A recent internet based survey of over 35,000 samples has shown that when different human observers are asked to assign labels to static human facial expressions, different individuals categorize differently the same image. This fact results in a lack of an unique ground-truth, an assumption held by the large majority of existing models for classification. This is especially true for highly ambiguous expressions, especially in the lack of a dynamic context. In this paper we propose to address this shortcoming by the use of discrete choice models (DCM) to describe the choice a human observer is faced to when assigning labels to static facial expressions. Different models of increasing complexity are specified to capture the causal effect between features of an image and its associated expression, using several combinations of different measurements. The sets of measurements we used are largely inspired by FACS but also introduce some new ideas, specific to a static framework. These models are calibrated using maximum likelihood techniques and they are compared with each other using a likelihood ratio test, in order to test for significance in the improvement resulting from adding supplemental features. Through a cross-validation procedure we assess the validity of our approach against overfitting and we provide a comparison with an alternative model based on Neural Networks for benchmark purposes.

Atlas-based Segmentation

Image segmentation is a main task in many medical applications such as surgical or radiation therapy planning, automatic labelling of anatomical structures or morphological and morphometrical studies. Segmentation in medical imaging is however challenging because of problems linked to low contrast images, fuzzy object-contours, similar intensities with adjacent objects of interest, etc. Using prior knowledge can help in the segmentation task. A widely used method consists to extract this prior knowledge from a reference image often called atlas. We review in this chapter the existing approaches for atlas-based segmentation in medical imaging and we focus on those based on a volume registration method. We present the problem of using atlas information for pathological image analysis and we propose our solution for atlas-based segmentation in MR image of the brain when large space-occupying lesions are present. Finally, we present the new research directions that aim at overcome current limitations of atlas-based segmentation approaches based on registration only.

Model-Based Segmentation and Fusion of 3D Computed Tomography and 3D Ultrasound of the Eye for Radiotherapy Planning

Computed Tomography (CT) represents the standard imaging modality for tumor volume delineation for radiotherapy treatment planning of retinoblastoma despite some inherent limitations. CT scan is very useful in providing information on physical density for dose calculation and morphological volumetric information but presents a low sensitivity in assessing the tumor viability. On the other hand, 3D ultrasound (US) allows a highly accurate definition of the tumor volume thanks to its high spatial resolution but it is not currently integrated in the treatment planning but used only for diagnosis and follow-up. Our ultimate goal is an automatic segmentation of gross tumor volume (GTV) in the 3D US, the segmentation of the organs at risk (OAR) in the CT and the registration of both modalities. In this paper, we present some preliminary results in this direction. We present 3D active contour-based segmentation of the eye ball and the lens in CT images; the presented approach incorporates the prior knowledge of the anatomy by using a 3D geometrical eye model. The automated segmentation results are validated by comparing with manual segmentations. Then, we present two approaches for the fusion of 3D CT and US images: (i) landmark-based transformation, and (ii) object-based transformation that makes use of eye ball contour information on CT and US images.

Spread spectrum for interferometric and magnetic resonance imaging

We consider images probed through incomplete and noisy Fourier coverages, both in the context of radio interferometry (RI) and of magnetic resonance imaging (MRI). We show that the quality of signal reconstruction can be significantly enhanced by the introduction of a linear chirp modulation, which induces a spread spectrum phenomenon.

EVALUATION OF BRAIN IMAGE NONRIGID REGISTRATION ALGORITHMS BASED ON LOG-EUCLIDEAN MR-DTI CONSISTENCY MEASURES

Several nonrigid registration algorithms have been proposed for inter-subject alignment, used to construct statistical atlases and to identify group differences. Assessment of the accuracy of nonrigid registration algorithms is an essential and complex issue due to its intricate framework and its application-dependent behavior. We demonstrate that the diffusion MRI provides an independent means of assessing the quality of alignment achieved on the structural MRI. Diffusion tensor MRI (DT-MRI) enables the comparison of the local position and orientation of regions that appear homogeneous in conventional MRI. We carried out inter-subject alignment of conventional T1-weighted MRI with three different registration algorithms. Consequently, we projected DT-MRI of each subject through the same inter-subject transformation. The quality of the inter-subject alignment is assessed by estimating the consistency of the aligned DT-MRI using the Log-Euclidean framework

Understanding the Genesis of Cardiac Signals in Electrical Impedance Tomography

Electrical impedance tomography (EIT) is a safe and low-cost imaging technology allowing the monitoring of ventilation. While most EIT studies have investigated respiration-related events, EIT-based cardiovascular applications have received increasing attention over the last years only. Variations in intra-thoracic blood volume induce impedance changes that can be monitored with EIT and used for the estimation of hemodynamic parameters. There is, however, increasing evidence that variations in blood volume are not the only factors contributing to cardiac impedance changes within the heart. The mechanical action of the myocardium and movement of the heart-lung interface are suspected to generate impedance changes of non-negligible amplitude. To test this hypothesis we designed a dynamic 2D bio-impedance model from segmented human magnetic resonance data. EIT simulations were performed and showed that EIT signals in the heart area might be dominated up to 70% by motion-induced impedance changes.

An efficient segmentation method for ultrasound images based on a semi-supervised approach and patch-based features

Segmenting ultrasound images is a challenging problem where standard unsupervised segmentation methods such as the well-known Chan-Vese method fail. We propose in this paper an efficient segmentation method for this class of images. Our proposed algorithm is based on a semi-supervised approach (user labels) and the use of image patches as data features. We also consider the Pearson distance between patches, which has been shown to be robust w.r.t speckle noise present in ultrasound images. Our results on phantom and clinical data show a very high similarity agreement with the ground truth provided by a medical expert.

Soil moisture estimation using synergy of optical, SAR and topographic data with Gaussian Process Regression

In this work we address the synergy of optical, SAR (Synthetic Aperture Radar) and topographic data in soil moisture retrieval over an Alpine area. As estimation technique, we consider Gaussian Process Regression (GPR). The test area is located in South Tyrol, Italy where the main land types are meadows and pastures. Time series of ASAR Wide Swath - SAR, optical, topographic and ancillary data (meteorological information and snow cover maps) acquired repetitively in 2010 were examined. Regarding optical data, we used both, daily MODIS reflectances, and daily NDVI, interpolated from the 16-day MODIS composite. Slope, elevation and aspect were extracted from a 2.5 m DEM (Digital Elevation Model) and resampled to 10 m. Daily soil moisture measurements were collected in the three fixed stations (two located in meadows and one located in pasture). The snow maps were used to mask the points covered by snow. The best performance was obtained by adding MODIS band 6 at 1640 nm to SAR and DEM features. The corresponding coefficient of determination, R2, was equal to 0.848, and the root mean square error, RMSE, to 5.4 % Vol. Compared to the case when no optical data were considered, there was an increase of ca. 0.05 in R2 and a decrease in RMSE of ca. 0.7 % Vol. This work showed that the joint use of NDVI or water absorption reflectance with SAR and topographic data can improve the estimation of soil moisture in specific Alpine area and that GPR is an effective method for estimation.