Back to Events

Seminar Series

July 09

Perfusion Imaging via Mass Transport

Peirong Liu
Martinos Center for Biomedical Imaging, MGB

Part Of

11:00A
- 12:00P

Location

32-370
Add to Calendar 2025-07-09 11:00:00 2025-07-09 12:00:00 America/New_York Perfusion Imaging via Mass Transport Perfusion imaging (PI) quantifies blood flow through the brain parenchyma by serial imaging. The resulting quantitative measures help clinical diagnosis and clinical decision-making, for example, to assess acute strokes and brain tumors. Despite its benefits, the widespread use of PI still faces many challenges: (1) Current existing perfusion analysis approaches mostly depend on the arterial input function (AIF), while the selection procedure for AIF is not unified and is only a coarse approximation of the actual input tracer; (2) These approaches are performed on individual voxels, thereby disregarding spatial dependencies of tracer dynamics. In this talk, I will present a series of our PDE-based approaches from a mass transport perspective, to better understand the relations between tracer’s spatial-temporal transport and strokes, meanwhile avoiding the need of approximating AIF. We (1) proposed a variable-coefficient advection-diffusion PDE framework, which models the tracer transport from both optimization- and learning-based perspectives; (2) introduced constraint-free representations of physically meaningful velocity and diffusion fields in the tracer’s PDE system. Looking forward, I am excited to explore advanced, physics-informed formulations for dynamic modeling in real-world clinical settings. A key focus will be the development of interactive models capable of real-time prediction of patient outcomes following interventional treatments. By bridging theoretical foundations and practical applications, my long-term vision is to strengthen the robustness of machine learning, address real-world challenges in clinical data, and ultimately contribute to a safer, more reliable, and accessible healthcare system.  TBD

June 06

Bayesian Deep Learning: From Reliable Neural Networks to Interpretable Foundation Models

Hao Wang
Rutgers University

Part Of

11:00A
- 12:00P

Location

32-D407
Add to Calendar 2025-06-06 11:00:00 2025-06-06 12:00:00 America/New_York Bayesian Deep Learning: From Reliable Neural Networks to Interpretable Foundation Models While perception tasks such as visual object recognition and text understanding play an important role in human intelligence, the subsequent tasks that involve inference, reasoning, and planning require an even higher level of intelligence.  The past decade has seen major advances in many perception tasks using deep learning models. In terms of higher-level inference, however, probabilistic graphical models, with their ability to expressively describe properties of variables and various probabilistic relations among variables, are still more powerful and flexible. To achieve integrated intelligence that involves both perception and inference, we have been exploring along a research direction, which we call Bayesian deep learning, to tightly integrate deep learning and Bayesian models within a principled probabilistic framework. In this talk, I will present the proposed unified framework and some of our recent work on Bayesian deep learning with various applications including interpretable large language models, network analysis, and healthcare. TBD

May 14

[Thesis Defense] Language-Centric Medical Image Understanding

Peiqi Wang
CSAIL, MIT

Part Of

3:00P
- 4:15P

Location

32-D463
Star
Add to Calendar 2025-05-14 15:00:00 2025-05-14 16:15:00 America/New_York [Thesis Defense] Language-Centric Medical Image Understanding This thesis improves how machines understand medical images by using language in three ways: as a source of supervision, as prior knowledge, and as a way to communicate results. The main contributions are:  (1) a method that uses the text in medical reports to help link specific parts of an image to the descriptions, (2) a way to make learning from noisy and imbalanced clinical data more robust by using language-based cues to correct bias on a case-by-case basis, and (3) a framework for assessing and improving how well linguistic expressions of diagnostic certainty are calibrated. Together, these contributions improve the accuracy, robustness, and reliability of medical AI systems, supporting more effective clinical workflows and improved patient care.  TBD

May 07

Using distance fields to represent and render shape

Sarah Frisken
Harvard Medical School

Part Of

11:00A
- 12:00P

Location

32-D407
Add to Calendar 2025-05-07 11:00:00 2025-05-07 12:00:00 America/New_York Using distance fields to represent and render shape Distance fields have advantages over classic graphical shape representations such as meshes and point clouds for pervasive graphics problems such as shape construction, collision detection, and anti-aliasing. I will describe distance field-based applications that we built over the years for 3D sculpting, font rendering, digital drawing, CNC milling simulation, and more recent work in medical image processing. I will focus on innovations that were required to address size, speed, and accuracy in these systems, share some stories about commercialization, and describe new research directions. TBD

April 09

SDF Geometry Processing

Silvia Sellán
CSAIL

Part Of

11:00A
- 12:00P

Location

32-370
Add to Calendar 2025-04-09 11:00:00 2025-04-09 12:00:00 America/New_York SDF Geometry Processing The historical focus of the Computer Graphics research community has had a deep influence on the representations chosen to store and process geometric information. Beyond the classical explicit formats like meshes and point clouds; in this talk, I will argue that Signed Distance Functions (SDFs) are a preferable representation for many tasks in engineering, robotics and machine learning. I will review recent work on reconstructing surfaces from SDFs and share an exciting vision for a future of geometric deep learning that exploits all the geometric information contained in each SDF sample. TBD

November 21

Geometric Algebra Planes: Convex Implicit Neural Volumes

Sara Fridovich-Keil
Stanford University

Part Of

11:00A
- 12:00P

Location

32-D507
Add to Calendar 2024-11-21 11:00:00 2024-11-21 12:00:00 America/New_York Geometric Algebra Planes: Convex Implicit Neural Volumes Volume parameterizations abound in recent literature, from the classic voxel grid to the implicit neural representation and everything in between. While implicit representations have shown impressive capacity and better memory efficiency compared to voxel grids, to date they require training via nonconvex optimization. This nonconvex training process can be slow to converge and sensitive to initialization and hyperparameter choices that affect the final converged result. We introduce a family of models, GA-Planes, that is the first class of implicit neural volume representations that can be trained by convex optimization. GA-Planes models include any combination of features stored in tensor basis elements, fol- lowed by a neural feature decoder. They generalize many existing representations and can be adapted for convex, semiconvex, or nonconvex training as needed for different inverse problems. In the 2D setting with a linear feature decoder, we prove that GA-Planes is equivalent to a low-rank plus low-resolution matrix factorization; we show that this approximation outperforms the classic low-rank plus sparse decomposition for fitting a natural image. In 3D, we demonstrate GA-Planes’ competitive performance in terms of expressiveness, model size, and optimizability across three volume fitting tasks: radiance field reconstruction, 3D segmentation, and video segmentation. 32-D507

October 24

Fetal MRI: motion correction, advanced modalities and clinical applications

Maria Deprez
King's College London

Part Of

11:00A
- 12:00P

Location

32-D451
Add to Calendar 2024-10-24 11:00:00 2024-10-24 12:00:00 America/New_York Fetal MRI: motion correction, advanced modalities and clinical applications Fetal MRI enables visualisation of developing fetal brain and organs, and it is increasingly used in both research and clinical practice. However, fetal motion and maternal breathing present challenges for acquisition and processing of fetal MRI data, which cannot be addressed by standard tools for adult or neonatal populations. This talk will describe dedicated correction techniques specifically suited for reconstructing artefact-free fetal MRI, including structural and diffusion imaging of the brain, and imaging of moving fetal heart. First, we will introduce the key concepts of correction of rigid and deformable motion, and super-resolution reconstruction of fetal structural MRI. Next, registration-based distortion correction will be introduced to correct geometric distortion present in echo-planar imaging used for acquisition of diffusion data, and super-resolution technique will be extended to reconstruct high angular resolution diffusion imaging (HARDI) of fetal brain. Finally, we will describe motion correction and super-resolution reconstruction technique to reconstruct images of moving fetal heart. For each of these modalities we will showcase example clinical applications and preliminary findings. 32-D451