September 18

Add to Calendar 2017-09-18 16:00:00 2017-09-18 17:00:00 America/New_York Exploring Uncertainty in dMRI Super-resolution via Probabilistic CNNs Deep learning has shown success in a wide range of medical image transformation problems, such as super-resolution (SR), denoising and image synthesis. However, the highly ill-posed nature of such problems results in inevitable ambiguity in the learning of networks. In addition, the deterministic nature of the existing methods means that they provide no indication of confidence in its prediction, which hinders reliability assessment and forms a significant barrier to adoption in clinical practice.This talks will focus on our recent work which proposes a probabilistic CNN method for modelling uncertainty in image enhancement problems. We propose to account for intrinsic uncertainty through a per-patch heteroscedastic noise model and for parameter uncertainty through approximate Bayesian inference in the form of variational dropout. We show that the combined benefits of both lead to the state-of-the-art performance super-resolution of diffusion MR brain images. We further show that the method produces tangible benefits in downstream tractography. In addition, the probabilistic nature of the methods naturally confers a mechanism to quantify uncertainty over the super-resolved output. We demonstrate through experiments on both healthy and pathological brains the potential utility of such an uncertainty measure in the risk assessment of the super-resolved images for subsequent clinical use. Conference Room D451 Belfer sarah_donahue@hks.harvard.edu

October 19

Add to Calendar 2017-10-19 16:00:00 2017-10-19 17:00:00 America/New_York Disease Progression Modelling of Alzheimer's Disease Subtypes Distinct neurodegenerative diseases such as typical Alzheimer'sdisease (AD) or Posterior Cortical Atrophy (PCA) exhibit differentdynamics of pathology progression in the brain. Understanding theprecise mechanisms of disease progression will enable us to createinformed drug treatments. Furthermore, quantifying the progression ofthese diseases using specialised mathematical models will enable us tostage subjects more accurately along the disease progression timeline,which is important for assessing drug efficacy. In this presentation Iwill talk about my work on developing novel disease progression modelsand their applications to typical AD and PCA. Finally, I willconclude with a few words on the TADPOLE challenge, which aims topredict the progression of individuals at risk of AD. 32-D507 Belfer sarah_donahue@hks.harvard.edu

December 07

Add to Calendar 2017-12-07 16:00:00 2017-12-07 17:00:00 America/New_York Computational analysis and modeling of graph-structured neuroimaging data Graph representations are often used to model structured data at anindividual or population level and have numerous applications inpattern recognition problems. In the field of neuroscience, where suchrepresentations are commonly used to model structural or functionalconnectivity between a set of brain regions, graphs have proven to beof great importance to reveal patterns related to brain developmentand disease, which were previously unknown. This talk is going tocover the evaluation of similarity between brain connectivity networksin a manner that accounts for the graph structure and is tailored fora particular application. At the same time, exploiting the wealth ofimaging and non-imaging information for disease prediction tasksrequires models capable of simultaneously representing individualfeatures and data associations between subjects from potentially largepopulations. The latter can be particularly beneficial in large-scalestudies and graphs provide a natural framework for suchtasks. Concepts from signal processing on graphs allow convolutions ona population graph incorporating both imaging and non-imaginginformation and this talk will demonstrate their importance forsemi-supervised classification tasks, inferring subject specificproperties from their imaging features and interactions within apopulation. 32-D507 Belfer sarah_donahue@hks.harvard.edu

March 05

Add to Calendar 2018-03-05 17:00:00 2018-03-05 18:00:00 America/New_York An Unsupervised Learning Model for Fast Deformable Medical Image Registration We present an efficient learning-based algorithm for deformable,pairwise 3D medical image registration. Current registration methodsoptimize an energy function independently for each pair of images,which can be time-consuming for large data. We define registration asa parametric function, and optimize its parameters given a set ofimages from a collection of interest. Given a new pair of scans, wecan quickly compute a registration field by directly evaluating thefunction using the learned parameters. We model this function using aCNN, and use a spatial transform layer to reconstruct one image fromanother while imposing smoothness constraints on the registrationfield. The proposed method does not require supervised informationsuch as ground truth registration fields or anatomical landmarks. Wedemonstrate registration accuracy comparable to state-of-the-art 3Dimage registration, while operating orders of magnitude faster inpractice. Our method promises to significantly speed up medical imageanalysis and processing pipelines, while facilitating novel directionsin learning-based registration and its applications.The pre-print is available at https://arxiv.org/abs/1802.02604 32-D407 Belfer sarah_donahue@hks.harvard.edu

March 12

Parallel RF Transmission for 3 Tesla and 7 Tesla MRI

Filiz Yetisir
Children's Hospital Boston; Harvard Medical School
Add to Calendar 2018-03-12 17:00:00 2018-03-12 18:00:00 America/New_York Parallel RF Transmission for 3 Tesla and 7 Tesla MRI High magnetic field strength (3 Tesla or 7 Tesla) MRI has become popular inthe last few decades due to the better signal to noise ratio and/orresolution it provides which leads to image quality improvement anddiagnostic advantages. Two challenges associated with high field MRI areincreased image shading and increased tissue heating which degrade theimage quality and limit imaging speed. Parallel transmission technology hasthe potential to address both the image shading and tissue heatingchallenges, however it is associated with implementation and safetychallenges. In this talk, the application of parallel transmission to brainimaging at 7 Tesla and fetal imaging at 3 Tesla MRI will be discussed. Athorough framework for safe and effective implementation of paralleltransmission is proposed and demonstrated through simulation studies aswell as phantom experiments. 32-D407 Belfer sarah_donahue@hks.harvard.edu

March 19

Add to Calendar 2018-03-19 17:00:00 2018-03-19 18:00:00 America/New_York MR relaxometry: Introduction and applications in the adult brain and placenta Though MRI is most often used clinically to perform structuralimaging, aspects of MR signals are sensitive tophysiology. Longitudinal (T1) and transverse (T2) relaxation timesdepend on many aspects of tissue microstructure, such as extracellularwater volume (T1), blood oxygen saturation (T1 and T2) and vascularstructure (T2 and T2*). Though relaxometry could inform clinicaldecision making, there are many challenges to acquiring reliablevalues for T1/T2/T2* in a clinical setting. For example, severalstrategies have been proposed for isolating a pure blood signal for T2analysis with the intention of determining blood oxygen saturation. Iwill present one of these techniques, called QUIXOTIC, to demonstratehow relaxometry can be used to quantify brain oxygen consumption andto illustrate some of the complications related to measuring therelaxivity of tissues. I will then describe ongoing efforts to usemagnetic resonance fingerprinting, a new approach to quantitative MRI,to determine the baseline physiology of the placenta. 32-D407 Belfer sarah_donahue@hks.harvard.edu

April 19

Add to Calendar 2018-04-19 16:00:00 2018-04-19 17:00:00 America/New_York Improving generalization of neural networks via unsupervised domain adaptation and semi-supervision Advances in machine learning (ML) have accelerated research in biomedicalimage analysis and promise to disrupt clinical practice by enabling tools forimproved diagnosis and treatment pipelines. However, the performance of MLsystems often degrades when they are applied on test data that differ from thetraining data, for example, due to variations in imaging quality and protocols.This hinders usage of ML systems on large-scale studies or outside the lab.The first part of this talk will introduce the issue of domain shift betweentraining and testing data distributions and the basic theory of domainadaptation that tries to alleviate it. It will then cover our work with adversarialnetworks for unsupervised domain adaptation, which investigated theirpotential for segmentation of lesions in multimodal MR images and proposedtheir extension with multi-connected networks for improved adaptation. It willthen provide insights in the behaviour and limitations of the approach.Motivated by a common limitation of unsupervised domain adaptationmethods, that they do not account for discriminative separation of theunlabelled target samples, the second part of the talk will focus on thecomplimentary problem of semi-supervision for a solution. After discussing thelow-density separation theorem I will present our recent work on regularizationof neural networks by capturing the manifold of unlabelled data in latent spacevia graphs and enforcing a network to learn more discriminativerepresentations. 32-D507 Belfer sarah_donahue@hks.harvard.edu

April 23

Add to Calendar 2018-04-23 17:00:00 2018-04-23 18:00:00 America/New_York A Statistical Imaging Framework for Magnetic Resonance Fingerprinting: Optimized Encoding and Decoding MRI scans are primarily performed and evaluated in a qualitative way using contrast-weighted images (e.g., with T1, T2 or proton-density weighting). These images weighting is a complex function of one or more of these intrinsic MR tissue parameters as modulated by external scanner settings and imperfections, providing limited capability for direct inter- and intra- patient comparisons across different institutions and/or across scanners. Although the potentials of quantitative MRI, which directly maps the underlying MR tissue parameters, have long been recognized, achieving this goal often requires lengthy acquisition times. Magnetic resonance fingerprinting (MRF) is a recent breakthrough in rapid quantitative MRI, which features a random or pseudo-random excitation scheme for encoding, as well as a dictionary matching scheme for decoding.Despite a brand new concept, the encoding and decoding processes for MRF are heuristic. In this talk, I will present our recent research that introduces a novel and rigorous statistical framework for optimized MRF. On the decoding side, I will introduce a principled statistical reconstruction approach, and show that the conventional MRF reconstruction is equivalent to the first iteration of the maximum likelihood reconstruction. On the encoding side, I will characterize the acquisition efficiency of MRF using estimation-theoretic bounds, and further use these bounds as metrics to perform optimal experimental design (e.g., designing flip angle and repetition time schedule). Surprisingly, the optimized acquisition parameters appear to be highly structured rather than randomly varying. In the end, I will make the connection between the optimal experiment design and optimal control theory, and discuss the structured behavior of optimized acquisition parameters. 32-D407 Belfer sarah_donahue@hks.harvard.edu

May 07

Add to Calendar 2018-05-07 17:00:00 2018-05-07 18:00:00 America/New_York Three-dimensional Cardiovascular Magnetic Resonance for Children with Congenital Heart Disease Severe congenital heart disease (CHD) affects approximately 1.2% of children and is the leading cause of birth defect-related deaths. Patients with severe CHD require multiple palliative surgeries during the first day of their lives to survive. Despite considerable improvement in the survival of these patients, there is increasing morbidity and mortality over time. It remains unclear why some of these patients fail their surgical repairs while others remain relatively well. Clinicians often rely on 2-dimensional (2D) images acquired from echocardiograms, catheterizations, or cardiovascular magnetic resonance (CMR) exams to assess these patients and qualitatively choose the optimal surgical repair. The 2D images, however, lead to a suboptimal understanding of the complex 3D spatial relationships and hemodynamics, and limit efficient decision making. To address this deficiency, we have developed a comprehensive 3D CMR sequence to acquire 3D block of images from patients with CHD. These 3D dataset is used for simultaneous assessment of cardiac function and hemodynamics. In this talk, we will introduce the conventional 2D CMR exam and discuss its limitations. The new 3D CMR exam will be then presented and compared with the 2D CMR exam in terms of cardiac function and hemodynamics. 32-D407 Belfer sarah_donahue@hks.harvard.edu

May 17

Add to Calendar 2018-05-17 17:00:00 2018-05-17 18:00:00 America/New_York Multidimensional diffusion imaging: Why we should leave conventional diffusion MRI behind The lecture will survey what features of tissue that can be probedwith diffusion MRI once we abandon conventional encoding techniques infavor of multidimensional diffusion encoding. The theory andmethodology of tensor-valued diffusion encoding will be studied indetail, and results will be demonstrated from healthy brain and braintumors. Time permitting, the methodology will be put in a largercontext together with related advanced imaging methods. 32-D407 Belfer sarah_donahue@hks.harvard.edu