535 research outputs found
Causality in digital medicine
No full textBen Glocker (an expert in machine learning for medical imaging, Imperial College London), Mirco Musolesi (a data science and digital health expert, University College London), Jonathan Richens (an expert in diagnostic machine learning models, Babylon Health) and Caroline Uhler (a computational biology expert, MIT) talked to Nature Communications about their research interests in causality inference and how this can provide a robust framework for digital medicine studies and their implementation, across different fields of application
Structure-guided registration in learning based image analysis
No full textImage registration is a key component in many medical image analysis pipelines and is useful in general computer vision applications. The goal of image registration is to find a transformation between the coordinate spaces of two images, such that the transformation aligns some Structure-of-Interest which exist in both images. Object tracking, image segmentation, multi-modal data fusion, longitudinal studies, label propagation, image labelling, population studies, image stitching and voxel based morphometry either rely on or at least benefit from image registration.
In this thesis, three aspects of image registration are discussed. Firstly, we utilise image registration to perform image segmentation via template deformation, the registration of some prior shape model with an image. We utilise neural networks to perform this template registration, the networks implicitly embed Structure-of-Interest information during training, to utilise this during inference when Structure-of-Interest information is not readily available. This differs from the conventional template deformation paradigm, where one must construct some image to segmentation likelihood function for the registration algorithm, a proxy function for the true segmentation accuracy. Utilising neural networks circumvents having to do this, we are able to train a network directly using a segmentation loss without hand crafting such a loss function. Our method gives us the prior enforcing benefits of template deformations without the difficulty of deriving some approximation to the segmentation loss.
Secondly, we develop a framework for combining iterative image registration with neural network based representation learning. Recent network based image registration has generally focused on improving the speed of registration, as neural networks are able to predict deformation fields in one shot, rather than iteratively converging during test time. We argue however that this comes at the cost of the accuracy of the registration. We propose a method that extracts out a feature representation that is well suited to be registered by any downstream registration algorithm. This exploits the ability of neural networks to discover rich representations of data and combining it with all the strengths of traditional registration algorithms.
Thirdly, we show why image registration is useful, we introduce an algorithm which extends Gaussian processes such that they can be successfully applied to large scale, 3D medical data. Gaussian processes are not generally well suited to image data, as computing covariances between image pixels is often nonsensical, unless however images are registered. By utilising registration, we are able to introduce patch kernels to allow for more anatomically dependent covariances between images.
Finally, we introduce two interesting new topics of research which we believe have relations and applications to image registration in a bid to further fuel discussion and investigation in the field.Open Acces
Application of the nonsmooth dynamics approach to model and analyze the contact-impact events in cam-follower systems
No full textThe dynamic modeling and analysis of planar rigid multibody systems that experience contact-impact events is presented and discussed throughout this work. The methodology is based on the nonsmooth dynamics approach, in which the interaction of the colliding bodies is modeled with multiple frictional unilateral constraints. Rigid multibody systems are stated as an equality of measures, which are formulated at the velocity-impulse level. The equations of motion are complemented with constitutive laws for the forces and impulses in the normal and tangential directions. In this work, the unilateral constraints are described by a set-valued force law of the type of Signorini’s condition, while the frictional contacts are characterized by a set-valued force law of the type of Coulomb’s law for dry friction. The resulting contact-impact problem is formulated and solved as an augmented Lagrangian approach, which is embedded in the Moreau time-stepping method. The effectiveness of the methodologies presented in this work is demonstrated throughout the dynamic simulation of a cam-follower system of an industrial cutting file machine.This work is supported by the Portuguese Foundation for the Science and Technology under the research project BIOJOINTS (PTDC/EME-PME/099764/2008). The first author expresses his gratitude to the Portuguese Foundation for the Science and Technology for the postdoctoral scholarship (SFRH/BPD/40067/2007). This research was conducted during a post-doctoral stay of the first author with Professor Christoph Glocker at the Center of Mechanics, ETH Zurich
Automatic Reporting of TBI Lesion Location in CT based on Deep Learning and Atlas Mapping
No full textTese de mestrado integrado, Engenharia Biomédica e Biofísica (Biofísica Médica e Fisiologia de Sistemas), 2021, Universidade de Lisboa, Faculdade de CiênciasThe assessment of Computed Tomography (CT) scans for Traumatic Brain Injury (TBI) management remains a time consuming and challenging task for physicians. Computational methods for quantitative lesion segmentation and localisation may increase consistency in diagnosis and prognosis criteria.
Our goal was to develop a registration-based tool to accurately localise several lesion classes (i.e., calculate the volume of lesion per brain region), as an extension of the Brain Lesion Analysis and Segmentation Tool for CT (BLAST-CT).
Lesions were located by projecting a Magnetic Resonance Imaging (MRI) labelled atlas from the
Montreal Neurological Institute (MNI MRI atlas) to a lesion map in native space. We created a CT
template to work as an intermediate step between the two imaging spaces, using 182 non-lesioned CT
scans and an unbiased iterative registration approach. We then non-linearly registered the parcellated
atlas to the CT template, subsequently registering (affine) the result to native space. From the final atlas
alignment, it was possible to calculate the volume of each lesion class per brain region. This pipeline
was validated on a multi-centre dataset (n=839 scans), and defined three methods to flag any scans that
presented sub-optimal results. The first one was based on the similarity metric of the registration of every
scan to the study-specific CT template, the second aimed to identify any scans with regions that were
completely collapsed post registration, and the final one identified scans with a significant volume of
intra-ventricular haemorrhage outside of the ventricles. Additionally, an assessment of lesion prevalence
and of the false negative and false positive rates of the algorithm, per anatomical region, was conducted,
along with a bias assessment of the BLAST-CT tool.
Our results show that the constructed pipeline is able to successfully localise TBI lesions across
the whole brain, although without voxel-wise accuracy. We found the error rates calculated for each
brain region to be inversely correlated with the lesion volume within that region. No considerable bias
was identified on BLAST-CT, as all the significant correlation coefficients calculated between the Dice
scores and clinical variables (i.e., age, Glasgow Coma Scale, Extended Glasgow Outcome Scale and
Injury Severity Score) were below 0.2. Our results also suggest that the variation in DSC between male
and female patients within a specific age range was caused by the discrepancy in lesion volume presented
by the scans included in each sample
Automatic Reporting of TBI Lesion Location in CT based on Deep Learning and Atlas Mapping
No full textTese de mestrado integrado, Engenharia Biomédica e Biofísica (Biofísica Médica e Fisiologia de Sistemas), 2021, Universidade de Lisboa, Faculdade de CiênciasThe assessment of Computed Tomography (CT) scans for Traumatic Brain Injury (TBI) management remains a time consuming and challenging task for physicians. Computational methods for quantitative lesion segmentation and localisation may increase consistency in diagnosis and prognosis criteria.
Our goal was to develop a registration-based tool to accurately localise several lesion classes (i.e., calculate the volume of lesion per brain region), as an extension of the Brain Lesion Analysis and Segmentation Tool for CT (BLAST-CT).
Lesions were located by projecting a Magnetic Resonance Imaging (MRI) labelled atlas from the
Montreal Neurological Institute (MNI MRI atlas) to a lesion map in native space. We created a CT
template to work as an intermediate step between the two imaging spaces, using 182 non-lesioned CT
scans and an unbiased iterative registration approach. We then non-linearly registered the parcellated
atlas to the CT template, subsequently registering (affine) the result to native space. From the final atlas
alignment, it was possible to calculate the volume of each lesion class per brain region. This pipeline
was validated on a multi-centre dataset (n=839 scans), and defined three methods to flag any scans that
presented sub-optimal results. The first one was based on the similarity metric of the registration of every
scan to the study-specific CT template, the second aimed to identify any scans with regions that were
completely collapsed post registration, and the final one identified scans with a significant volume of
intra-ventricular haemorrhage outside of the ventricles. Additionally, an assessment of lesion prevalence
and of the false negative and false positive rates of the algorithm, per anatomical region, was conducted,
along with a bias assessment of the BLAST-CT tool.
Our results show that the constructed pipeline is able to successfully localise TBI lesions across
the whole brain, although without voxel-wise accuracy. We found the error rates calculated for each
brain region to be inversely correlated with the lesion volume within that region. No considerable bias
was identified on BLAST-CT, as all the significant correlation coefficients calculated between the Dice
scores and clinical variables (i.e., age, Glasgow Coma Scale, Extended Glasgow Outcome Scale and
Injury Severity Score) were below 0.2. Our results also suggest that the variation in DSC between male
and female patients within a specific age range was caused by the discrepancy in lesion volume presented
by the scans included in each sample
Efficient extraction of semantic information from medical images in large datasets using random forests
No full textLarge datasets of unlabelled medical images are increasingly becoming available; however only a small subset tend to be manually semantically labelled as it is a tedious and extremely time-consuming task to do for large datasets.
This thesis aims to tackle the problem of efficiently extracting semantic information in the form of image segmentations and organ localisations from large datasets of unlabelled medical images. To do so, we investigate the suitability of supervoxels and random classification forests for the task.
The first contribution of this thesis is a novel method for efficiently estimating coarse correspondences between pairs of images that can handle difficult cases that exhibit large variations in fields of view. The proposed methods adapts the random forest framework, which is a supervised learning algorithm, to work in an unsupervised manner by automatically generating labels for training via the use of supervoxels.
The second contribution of this thesis is a method that extends our first contribution so as to be applicable efficiently on a large dataset of images. The proposed method is efficient and can be used to obtain correspondences between a large number of object-like supervoxels that are representative of organ structures in the images. The method is evaluated for the applications of organ-based image retrieval and weakly-supervised image segmentation using extremely minimal user input. While the method does not achieve image segmentation accuracies for all organs in an abdominal CT dataset compared to current fully-supervised state-of-the-art methods, it does provide a promising way for efficiently extracting and parsing a large dataset of medical images for the purpose of further processing.Open Acces
Polarized light field Imaging for reflectance separation and transparent object acquisition
No full textThis thesis addresses two challenging acquisition problems for computer graphics: acquisition of layered reflectance, and acquiring common transparent objects (axially-symmetric containers) and liquids. The first problem is approached using a novel computational photography method for single shot separation of diffuse/specular reflectance as well as novel angular domain separation of layered reflectance. A polarized light field camera is presented as a key component for simultaneously capturing multiple states of polarization and this thesis demonstrates single-shot reflectance separation under various types of incident illumination. To tackle the second problem, a practical solution is proposed for high quality reconstruction of axially-symmetric transparent objects and estimation of the optical properties of various common liquids contained in such objects. Acquisition setup of the solution involves imaging such objects from a single viewpoint while illuminating them from directly behind with polarized patterns. A reconstruction step is then based on optimization of the object's geometry and its refractive index to minimize the difference between observed and simulated transmission and refraction of rays passing through the object. The same setup is also employed for developing a novel method for estimating optical properties of common liquids contained in natural concentrations in a known axially-symmetric transparent container.Open Acces
Understanding and mitigating universal adversarial perturbations for computer vision neural networks
No full textDeep neural networks (DNNs) have become the algorithm of choice for many computer vision applications. They are able to achieve human level performance in many computer vision tasks, and enable the automation and large-scale deployment of applications such as object tracking, autonomous vehicles, and medical imaging. However, DNNs expose software applications to systemic vulnerabilities in the form of Universal Adversarial Perturbations (UAPs): input perturbation attacks that can cause DNNs to make classification errors on large sets of inputs.
Our aim is to improve the robustness of computer vision DNNs to UAPs without sacrificing the models' predictive performance. To this end, we increase our understanding of these vulnerabilities by investigating the visual structures and patterns commonly appearing in UAPs. We demonstrate the efficacy and pervasiveness of UAPs by showing how Procedural Noise patterns can be used to generate efficient zero-knowledge attacks for different computer vision models and tasks at minimal cost to the attacker. We then evaluate the UAP robustness of various shape and texture-biased models, and found that applying them in ensembles provides marginal improvement to robustness.
To mitigate UAP attacks, we develop two novel approaches. First, we propose the Jacobian of DNNs to measure the sensitivity of computer vision DNNs. We derive theoretical bounds and provide empirical evidence that shows how a combination of Jacobian regularisation and ensemble methods allow for increased model robustness against UAPs without degrading the predictive performance of computer vision DNNs. Our results evince a robustness-accuracy trade-off against UAPs that is better than those of models trained in conventional ways. Finally, we design a detection method that analyses the hidden layer activation values to identify a variety of UAP attacks in real-time with low-latency. We show that our work outperforms existing defences under realistic time and computation constraints.Open Acces
- …
