Search

Unsupervised dimensionality reduction (DR) aims to preserve input data structure in a low-dimensional (LD) space based on neighborhood information. In contrast, supervised DR intends to improve the learning performance, i.e., classification and regression, in an LD representation. Unfortunately, obtaining the complete label outputs of a data set for real-world applications is hard. Here, we introduce a novel DR framework coupling both available class labels and input feature similarities to extend the well-known t-distributed Stochastic Neighbor Embedding (SNE) for semi-supervised scenarios. Our proposal, termed Semi-Supervised t-SNE (SS.t-SNE), properly fixes the widths of Gaussian neighborhoods to reveal the salient local and global data structures in an LD space. Indeed, our approach is presented as a generalization of unsupervised and supervised versions of t-SNE. SS.t-SNE outperforms other semi-supervised DR methods in data visualization and classification tasks in LD embeddings.

Noisy multi-way data sets are ubiquitous in many domains. In neuroscience, electroencephalogram (EEG) data are recorded during periodic stimulation from different sensory modalities, leading to steady-state (SS) recordings with at least four ways: the channels, the time, the subjects and the modalities. Improving the signal-to-noise ratio (SNR) of the SS responses is crucial to enable their practical use. Supervised spatial filtering methods can be considered for this purpose to relevantly guide the extraction of specific activity patterns. Nevertheless, such approaches are difficult to validate with few subjects and can process at most two data ways simultaneously, the remaining ones being either averaged or considered independently despite their dependencies. This paper hence designs unsupervised tensor factorization models to enable identifying meaningful underlying structures characterized in all ways of multimodal SS data. We show on EEG recordings from 15 subjects that such factorizations faithfully reveal consistent spatial topographies, time courses with enhanced SNR and subject variations of the periodic brain activity.

Stochastic Neighbor Embedding (SNE) and variants like t-distributed SNE are popular methods of unsupervised dimensionality reduction (DR) that deliver outstanding experimental results. Regular t-SNE is often used to visualize data with class labels in colored scatterplots, even if those labels are actually not involved in the DR process. This paper proposes a modification of t-SNE that employs class labels to adjust the widths of the Gaussian neighborhoods around each datum, instead of deriving those from a perplexity set by the user. The widths are fixed to concentrate a major fraction of the probability distribution around a datum on neighbors with the same class. This tends to shrink the bulk of the classes and to stretch their low-dimensional separation. Experimental results show that the proposed class-aware t-SNE (cat-SNE) outperforms regular t-SNE in KNN classification tasks carried out in the embedding.

For prostate cancer patients, large organ deformations occurring between the sessions of a fractionated radiotherapytreatment lead to uncertainties in the doses delivered to the tumour and the surrounding organs at risk. Thesegmentation of those structures in cone beam CT (CBCT) volumes acquired before every treatment sessionis desired to reduce those uncertainties. In this work, we perform a fully automatic bladder segmentation ofCBCT volumes with u-net, a 3D fully convolutional neural network (FCN). Since annotations are hard to collectfor CBCT volumes, we consider augmenting the training dataset with annotated CT volumes and show that itimproves the segmentation performance.Our network is trained and tested on 48 annotated CBCT volumes using a 6-fold cross-validation scheme.The network reaches a mean Dice similarity coefficient (DSC) of0.801±0.137 with 32 training CBCT volumes.This result improves to0.848±0.085 when the training set is augmented with 64 CT volumes. The segmentationaccuracy increases both with the number of CBCT and CT volumes in the training set. As a comparison, thestate-of-the-art deformable image registration (DIR) contour propagation between planning CT and daily CBCTavailable in RayStation reaches a DSC of0.744±0.144 on the same dataset, which is below our FCN result.

International audience; Freeze-thaw (FT) stability is often used to assess the ability of a gel to support the damage induced byfreezing; selected parameters such as drip loss, damage to structure etc can be used to assess the freezetolerance of a gel. Konjac glucomannan (KGM) is a very specific hydrocolloid able to trap 100 timesits weight in water; it has not been studied so far as an improver to enhance FT stability. The aim of the study was to show that the presence of a small quantity of konjac glucomannan (KGM)in potato starch suspension increased the stability of carvacrol antioxidant trapping. FT cycles wereused to accelerate the ageing of the product and to assess its stability. In addition to drip lossesdetermination, the stability of carvacrol trapping was evaluated by the quantification of carvacrol inthe syneresis liquid. Microscopic and macroscopic scales were considered with microscopy. Themoment of the addition of carvacrol and the presence of KGM both had an effect on the stability ofcarvacrol trapping and of the structure of the gel. KGM promoted amylose retrogradation but sloweddown amylopectin retrogradation. The stability of potato starch gels can be improved by the additionof a small quantity of KGM, which showed a “cryoprotectant” behaviour. New method to characterizethe micro and macrostructure from SEM images processing has also been proposed. The processing ofmicroscopy images was done using Generalized Fourier Descriptors and allowed the characterizationof each sample. The carvacrol addition lowered the physical stability of the gel with larger pores andincreased syneresis. On the contrary, the KGM addition increased the size of the pores but preventedthe formation of very large pores and reduced syneresis. The most stable system was obtained by theaddition of carvacrol at the end of heating, in a konjac glucomannane potato starch gel.

The lensless endoscope is a promising device designed to image tissues in vivo at the cellular scale. The traditional acquisition setup consists in raster scanning during which the focused light beam from the optical fiber illuminates sequentially each pixel of the field of view (FOV). The calibration step to focus the beam and the sampling scheme both take time. In this preliminary work, we propose a scanning method based on compressive sampling theory. The method does not rely on a focused beam but rather on the random illumination patterns generated by the single-mode fibers. Experiments are performed on synthetic data for different compression rates (from 10 to 100% of the FOV).

Stochastic Neighbor Embedding (SNE) and variants are dimensionality reduction (DR) methods able to foil the curse of dimensionality to deliver outstanding experimental results. Mitigating the crowding problem, t-SNE became an extremely popular DR scheme. Its quadratic time complexity in the number of samples is nevertheless unaffordable for big data sets. This motivates its Barnes-Hut (BH) acceleration for large-scale use. Although the latter is faster by orders of magnitude, few studies quantify its DR quality with respect to t-SNE. Extensive comparisons between t-SNE and its BH version are conducted using neighborhood preservation-based criteria. Both methods perform very similarly, suggesting the BH scheme superiority thanks to its reduced time complexity.

With the advent of Deep Learning (DL) techniques, especially Generative Adversarial Networks (GANs), data augmentation and genera- tion are quickly evolving domains that have raised much interest recently. However, the DL techniques are data demanding and since, medical data is not easily accessible, they suffer from data insufficiency. To deal with this limitation, different data augmentation techniques are used. Here, we propose a novel unsupervised data-driven approach for data augmentation that can generate 2D Computed Tomography (CT) images using a simple GAN. The generated CT images have good global and local features of a real CT image and can be used to augment the training datasets for effective learning. In this proof-of-concept study, we show that our pro- posed solution using GANs is able to capture some of the global and local CT variabilities. Our network is able to generate visually realistic CT im- ages and we aim to further enhance its output by scaling it to a higher resolution and potentially from 2D to 3D.

In dimensionality reduction and data visualisation, t-SNE has become a popular method. In this paper, we propose two variants to the Gaussian similarities used to characterise the neighbourhoods around each high-dimensional datum in t-SNE. A first alternative is to use t distributions like already used in the low-dimensional embedding space; a variable degree of freedom accounts for the intrinsic dimensionality of data. The second variant relies on compounds of Gaussian neighbourhoods with growing widths, thereby suppressing the need for the user to adjust a single size or perplexity. In both cases, heavy-tailed distributions thus characterise the neighbourhood relationships in the data space. Experiments show that both variants are competitive with t-SNE, at no extra cost.

Automatic delineation of organs at risk (OAR) in Computed Tomography (CT) images is a crucial step for treatment planning in radiation oncology. However, manual delineation of organs is a challenging and time-consuming task subject to inter-observer variabilities. Automatic organ delineation has been relying on non-rigid registrations and atlases. However, lately deep learning appears as a strong competitor with specific architectures dedicated to image segmentation like UNet. In this paper, we first assess the standard UNet to delineate multiple organs in CT images. Second, we observe the effect of dilated convolutional layers in UNet to better capture the global context from the CT images and effectively learn the anatomy, which results in increased accuracy of organ delineation. We evaluate the performance of a standard UNet and a dilated UNet (with dilated convolutional layers) on four chest organs (esophagus, left lung, right lung, and spinal cord) from 29 lung image acquisitions and observe that dilated UNet delineates the soft tissues notably esophagus and spinal cord with higher accuracy than the standard UNet. We quantify the segmentation accuracy of both models by computing spatial overlap measures like Dice similarity coefficient, recall & precision, and Hausdorff distance. Compared to the standard UNet, dilated UNet yields the best Dice scores for soft organs whereas for lungs, both models have the same delineation accuracy: 0.84±0.07 vs 0.71±0.10 for esophagus, 0.99±0.01 vs 0.99±0.01 for left lung, 0.99±0.01 vs 0.99±0.01 for right lung and 0.91±0.05 vs 0.88±0.04 for spinal cord.

The goal of visual analytics (VA) systems is to solve complex problems by integrating automated data analysis methods, such as machine learning (ML) algorithms, with interactive visualizations. We propose a conceptual framework that models human interactions with ML components in the VA process, and makes the crucial interplay between automated algorithms and interactive visualizations more concrete. The framework is illustrated through several examples. We derive three open research challenges at the intersection of ML and visualization research that will lead to more effective data analysis.

Images from positron emission tomography (PET) provide metabolic information about the human body. They present, however, a spatial resolution that is limited by physical and instrumental factors often modeled by a blurring function. Since this function is typically unknown, blind deconvolution (BD) techniques are needed in order to produce a useful restored PET image. In this work, we propose a general BD technique that restores a low resolution blurry image using information from data acquired with a high resolution modality (e.g., CT-based delineation of regions with uniform activity in PET images). The proposed BD method is validated on synthetic and actual phantoms. Images from positron emission tomography (PET) provide metabolic information about the human body. They present, however, a spatial resolution that is limited by physical and instrumental factors often modeled by a blurring function. Since this function is typically unknown, blind deconvolution (BD) techniques are needed in order to produce a useful restored PET image. In this work, we propose a general BD technique that restores a low resolution blurry image using information from data acquired with a high resolution modality (e.g., CT-based delineation of regions with uniform activity in PET images). The proposed BD method is validated on synthetic and actual phantoms.

Positron emission tomography is more and more used in radiation oncology, since it conveys useful functional information about cancerous lesions. Its rather low spatial resolution, however, prevents accurate tumor delineation and heterogeneity assessment. Post-reconstruction deconvolution with the measured point-spread function can address this issue, provided it does not introduce undesired artifacts. These usually result from inappropriate regularization, which is either absent or making too strong assumptions about the structure of the signal. This paper proposes a deconvolution method that is based on inverse problem theory and involves a new regularization term that preserves local pixel value order relationships. Such regularization entails relatively mild constraints that are directly inferred from the observed data. This paper investigates the theoretical properties of the proposed regularization and describes its numerical implementation with a primal-dual algorithm. Preliminary experiments with synthetic images are presented to compare quantitatively and qualitatively the proposed method to other regularization schemes, like TV and TGV. Positron emission tomography is more and more used in radiation oncology, since it conveys useful functional information about cancerous lesions. Its rather low spatial resolution, however, prevents accurate tumor delineation and heterogeneity assessment. Post-reconstruction deconvolution with the measured point-spread function can address this issue, provided it does not introduce undesired artifacts. These usually result from inappropriate regularization, which is either absent or making too strong assumptions about the structure of the signal. This paper proposes a deconvolution method that is based on inverse problem theory and involves a new regularization term that preserves local pixel value order relationships. Such regularization entails relatively mild constraints that are directly inferred from the observed data. This paper investigates the theoretical properties of the proposed regularization and describes its numerical implementation with a primal-dual algorithm. Preliminary experiments with synthetic images are presented to compare quantitatively and qualitatively the proposed method to other regularization schemes, like TV and TGV.

This work presents an approach allowing for an interactive visualization of dimensionality reduction outcomes, which is based on an extended view of conventional homotopy. The pairwise functional followed from a simple ho- motopic function can be incorporated within a geometrical framework in order to yield a bi-parametric approach able to combine several kernel matrices. There- fore, the users can establish the mixture of kernels in an intuitive fashion by only varying two parameters. Our approach is tested by using kernel alternatives for conventional methods of spectral dimensional reduction such as multidimensional scalling, locally linear embedding and laplacian eigenmaps. Provided mixture rep- resents every single dimensional reduction approach as well as helps users to find a suitable representation of embedded data.

Radiation therapy is one of the main treatments for cancer care. It consists in irradiating the tumor, while limiting the toxicity associated with the exposure of healthy tissues. Proton therapy is an emerging radiation delivery modality, which has the potential to better spare healthy tissues than conventional radiotherapy treatments. However, this new modality is much more sensitive to treatment uncertainties, such as patient anatomy changes. To address the lack of robustness in proton therapy, this thesis provides accurate treatment preparation tools, such as fast Monte Carlo dose calculation and robust planning methods. Furthermore, a comprehensive and realistic treatment robustness verification tool was developed in order to assess the sensitivity of the treatment plan to uncertainties. By combining these tools, proton therapy could be delivered more safely, improving the treatment outcome for the patient. All tools developed during this thesis are released open source and are already used in several institutions for research and clinical purposes. (BIFA - Sciences biomédicales et pharmaceutiques) -- UCL, 2018

University rankings play nowadays a major role in the decision of many students with regards to their future schools. Nonetheless, these rankings often remain quite opaque: not all data are made available, the methodology behind the rankings is not well defined, etc. One of the main ranking centred on business schools is the "Master in Management" from the Financial Times. This work aims to study the relevance of this ranking and its possible flaws. Several techniques are conducted, as a robustness analysis to assess the sturdiness of the ranking when facing uncertainties or dimensionality reductions to visualize the data set in two or three dimensions and thus envision local neighbourhoods of schools. Finally, from these analyses, potential improvements to the ranking from the Financial Times are discussed, from the design of a new ranking from scratch to the addition of visualization tools to enhance the informative character of the ranking. Master [120] : ingénieur civil en informatique, Université catholique de Louvain, 2017

A lot of applications on smartphones and other wearables use a localization system to trace the users. We will be interested in giving meaning to GPS traces (latitude / longitude / timestamp) by detecting meaningful locations for a user from the traces. The purpose of this master thesis is to find out an algorithm and a set of input parameters that gives the most satisfactory results to get the meaningful locations from datasets. This master thesis was realized in collaboration with SONY. We build up a very first intuitive solution from scratch, without any help of specialized literature. We enlarge a public dataset with calculation of speeds between successive observations, to only conserve the observations with a speed lower than 2 km/h. We display observations on maps with the tools of Google, but the solution is limited to public datasets. We also try to get our first meaningful locations with Excel sheets by grouping other observations with a chosen radius; it could be a solution, but this one is not scalable : each observation has successively to be compared to all other ones. The concept of "density" will then be used. The unsupervised learning machine domain will help to determinate clusters from datasets with algorithms, by grouping data based on similarity. DBSCAN is such an algorithm that defines a cluster as a maximal set of density-connected points, based on the "-neighborhood and the minimum of points per cluster. OPTICS is a more advance version of DBSCAN : the density of a cluster can be different from one cluster to another. K-MEANS is another type of algorithm which creates K clusters to minimize a function : the squared Euclidean distance from a point to the mean of the points to its cluster. With OPTICS in combination with K-MEANS, we build a program to try the several combinations of the algorithms with parameters sets (input values) on the five datasets from SONY. We get the results sets and do some validations to only keep realistic results. We switch then into supervised learning by using the ground truths as input for the validations. We assign a score depending on time - percentage of common duration in the clusters of the ground truths - and distance criteria - proximity of the centers of the clusters with the ground truths. The final score is the multiplication of both percentages. Each center of cluster should be a meaningful location. The highest score is held by this combination : run OPTICS and then reapply OPTICS on the same data without the outliers. This highest score is assigned to only one dataset, the other datasets have a really lower score. We then apply the geometric mean on the results of the highest score by the user for each combination of algorithms. The algorithm is OPTICS followed by K-MEANS : the centers of the clusters are determined by K-MEANS, based on the number of clusters determined by OPTICS. The most appropriated parameters from proposed parameter set are also determined. Master [60] en sciences informatiques, Université catholique de Louvain, 2017

The survival statistics for cancer patients treated with radiation therapy using photon beams show that many treatments fail due to poor tumour local control (TLC). A potential solution would be to increase the target dose, but this often entails high toxicity in the nearby healthy tissue. Unlike photons, protons release most of their energy at the end of their path (the so-called Bragg peak), reducing the dose to healthy tissue, which might be the key for safe dose escalation. Moreover, functional images can reveal spatial heterogeneity in the tumor radioresistance pattern and therefore allow for improved targeting (dose painting) and possibly enhanced TLC. In this context, special care must be taken to accurately model the possible uncertainties in the position of the Bragg peak, since they can strongly deteriorate treatment quality, especially in very heterogeneous dose distributions as in dose painting plans. In this thesis, we investigated the use of the most advanced techniques, such as Monte Carlo dose calculation and robust optimization, to ensure accurate and robust planning for (dose painted) proton therapy treatments. (BIFA - Sciences biomédicales et pharmaceutiques) -- UCL, 2017

External radiotherapy is a major component of cancer treatment. Its objective is to deliver a high measure of radiation to tumors while minimizing the dose inflicted to healthy tissues. Protons have the ability to deliver a peak of energy (the "Bragg peak") in a precise location depending on the atomic composition and densities of the tissues in their path, making them highly suitable for radiotherapy. Because the anatomy of the patient is continuously changing during the course of the treatment that may last more than 30 days, the position of the Bragg peak is subject to uncertainties, which may have dramatic consequences for the patient. The goal of this master thesis is to tune the treatment to day-to-day anatomical changes by performing the so-called "adaptive planning". We submit an original strategy where the treatment plan, optimized before the start of the treatment and approved by a radiation oncologist, is adjusted to the new anatomy. Three test cases have been designed: 1) anatomical changes around a tumor with unmodified position and shape; 2) tumor displacement; and 3) tumor reduction. The evaluation of a plan is conducted by analysing the dose volume histogram of the predicted dose by a Monte Carlo algorithm. Except for the tumor reduction, our adapted plans manage to produce a dose close to the desired one. In conclusion, non-rigid deformation of treatment plan could be used in addition of other method to perform efficient adaptive protontherapy. Master [120] : ingénieur civil biomédical, Université catholique de Louvain, 2016