Category: Senza categoria (Page 4 of 4)

Marco Palombo

Abstract

This talk will provide an introduction to microstructure imaging through dMRI and an overview of how machine learning can help improving several important aspects of it Diffusion MRI (dMRI) signal is sensitive to the tissue architecture at the cellular scale, namely microstructure. By analysing the dMRI measurements acquired with different experimental parameters, it is possible to infer some of the tissue properties (e.g. cell density, size and shape), with the ultimate goal of providing quantitative maps of tissue features and potentially define more specific biomarkers of the tissue state Conventional methods for microstructure feature mapping are based on fitting mathematical models to measured dMRI data. The fitting procedure have several limitations: it is time consuming, prone to degeneracy and require rich datasets Modern machine learning techniques can be used to learn the mapping between acquired dMRI signal and specific features of the tissue microstructure, by overcoming several of the major limitations of conventional model-fitting approach.

Short Bio

I got BSc. and MSc. degrees in Physics from Sapienza University of Rome, in Rome (Italy), on the broad topic of physics of soft matter and phase transitions Then I undertook a PhD in Biophysics always at Sapienza University to move my focus more on biological tissues, a special class of soft matter systems, and to study how to translate and apply concepts from soft matter physics to biology with the aim of improving the characterisation of biological tissue microstructure using non-invasive magnetic resonance imaging and spectroscopy techniques. My PhD work mostly focused on the characterisation of the microstructure of healthy brain tissue and brain tumor using diffusion weighted MRI For my first postdoc, I joined the Molecular Research Imaging Centre (MIRCen) in Paris (France) to expand my knowledge on molecular biology and neuroscience. I focused on developing metabolite-based diffusion MR spectroscopy techniques to quantify the structure of specific cell-types in the brain (e.g. neurons and glia) non-invasively and in-vivo. I pioneered a new paradigm that employs complex ultra-realistic virtual simulations of the brain cells to process the MR data using machine learning techniques. To further develop this innovative paradigm and apply it to broader MRI and the characterisation of organs other than the brain, I then took a position as senior research associate at the Centre for Medical Image Computing and Department of Computer Science at the University College London (UK) and most recently I was awarded the prestigious Future Leaders Fellowship.

Keywords

Microstructure imaging, Deep learning, Diffusion MRI, Machine learning, Regression

Giovanna Maria Dimitri

Abstract

In this talk I will briefly sketch applications of Deep Learning techniques to Brain MRI segmentation and reconstruction. The talk will cover some introductory aspects of Deep Learning models for segmentation, with a focus on Convolutional Neural Networks.

Short Bio

Dr Giovanna Maria Dimitri holds a PhD in Computer Science, obtained at the University of Cambridge (UK), under the supervision of Prof Pietro Lio’, with the dissertation: “Multilayer network methodologies for brain data analysis and modelling”. She graduated in July 2015 in the MPhil in Advanced Computer Science at the University of Cambridge, with distinction,under the supervision of Prof Pietro Liò with a dissertation on “Predicting Drugs Side Effects from a Combination of Chemical and Biological profiles”. She is a life member of Clare Hall college, Cambridge. Previously she received her master and bachelor thesis (both 110/110 cum laude) in Computer and Automation Engineering at the University of Siena. Her research interests are in developing deep learning and machine learning models for biomedical data analysis. She is currently a post-doc at the Department of Engineering at the University of Siena, working on deep learning for biomedical image processing. She is the author of several papers in international peer reviewed journals.

Keywords

Deep Learning, Brain MRI reconstruction, Brain MRI segmentation, Convolutional Neural Networks, Semantic Segmentation

Chen Qin

Abstract

Recent advances in deep learning have shown great potentials in improving the entire medical imaging pipeline, from image acquisition and reconstruction to disease diagnosis. In this talk, I will mainly focus on discussing deep learning for Magnetic Resonance (MR) image reconstruction. I will introduce our recent study on dynamic MR image reconstruction from highly undersampled k-space data, including the use of recurrent neural networks for modelling the sequential processes as well as exploiting complementary time and frequency domain knowledge for dynamic MRI reconstruction in both single-coil and multi-coil settings We show that deep learning is effective for dynamic MR image reconstruction in terms of both reconstruction quality and speed.

Short bio

Dr Chen Qin is a Lecturer in Computer Vision and Machine Learning at Electronics and Electrical Engineering, The University of Edinburgh. Before that, she worked as a Research Associate at Department of Computing, Imperial College London. She obtained her Ph.D. in Computing Research from Imperial College London. Her research is at the interdisciplinary field of artificial intelligence and medical imaging, aiming to improve the entire medical imaging/radiology workflow with significant impact for clinical use via machine intelligence. Her current research mainly focuses on the development of machine learning algorithms for magnetic resonance image reconstruction and analysis, including dynamic MR image reconstruction, medical image registration and segmentation.

Keywords

Deep Learning, Dynamic Magnetic Resonance Imaging, Cardiac Image Reconstruction, Recurrent Neural Networks

Mike Germuska

Abstract

Artificial neural networks are known to be robust estimators in the presence of noise. In this talk I show the potential for machine learning in physiological MRI for mapping cerebral oxygen metabolism. The performance of ML is compared to alternative statistical methods and the practicalities of applying ML to time series data is discussed.

Keywords

CMRO2, ANN, Machine Learning, Feature Selection, fMRI

Guy Gaziv

Abstract

Reconstructing seen natural images and decoding their novel semantic category from a subject’s evoked fMRI response is a milestone for developing brain-machine interfaces and for the study of consciousness. Unfortunately, acquiring sufficient (Image,fMRI) “paired” training data to span the huge space of natural images and their semantic classes is prohibitive, resulting in limited generalization power of today’s decoders. We present a novel self-supervised approach that overcomes the inherent lack of training data, simultaneously for both tasks — image reconstruction and large-scale semantic classification. Specifically, we impose cycle-consistency using two networks, encoder (E) & decoder (D), and train on additional “unpaired” data from the image and the fMRI domains. Concatenating those two networks back-to-back, E-D, allows for unsupervised training on unpaired images (i.e., images without fMRI recordings) — 50,000 natural images from 1000 ImageNet semantic categories in our experiments. Such self-supervision adapts the network to the statistics of novel images and their diverse categories. Similarly, concatenating our two networks, D-E, allows for unsupervised training on additional unpaired fMRI samples (i.e., fMRI recordings without images). Moreover, combining high-level perceptual reconstruction criteria with self-supervision on unpaired images results in a leap improvement over top existing methods, achieving unprecedented image-reconstruction from fMRI of never-before-seen images (evaluated by image metrics and human testing), and large-scale semantic classification (1000 diverse classes) of categories that are never-before-seen during network training (exceeding chance level accuracy by more than 100-fold). We further visualize the receptive field underlying our decoder and show the emergence of classic retinotopic organization. These results support the biological plausibility of our model.

Short bio

Guy Gaziv is a Computer Science PhD student at The Weizmann Institute of Science in the Michal Irani Computer Vision group. He previously studied human body motion during natural conversation at the Uri Alon lab (dept. of Molecular Cell Biology). He also worked as a Deep Learning researcher and a SW/HW engineer in the industry, including at Intel, Mellanox, FST biometrics, and the Israeli Intelligence Corps. His PhD research focuses on the intersection between machine and human vision, and specifically on decoding visual experience from brain activity. Guy earned his BSc in Electrical and Computer Engineering from The Hebrew University of Jerusalem and his MSc in Physics from The Weizmann Institute of Science.

Keywords

Deep Learning, Image Reconstruction, fMRI, Decoding, Reconstruction