Skip to main content
Add to Calendar
THESIS DEFENSE: Effective Modeling in Medical Imaging with Constrained Data
[ZOOM]https://mit.zoom.us/j/5323046147[TITLE]Effective Modeling in Medical Imaging with Constrained Data[SPEAKER]Tzu Ming Harry Hsu[ROOM]32-D463 (Star)[TIME]Aug 15, 15:00 - 16:00[ABSTRACT]Data for modern medical imaging modeling is constrained by their high physical density, complex structure, insufficient annotation, heterogeneity across sites, long-tailed distribution, and sparsely presented information. In this talk, I will cover three of my research to combat these scenarios: the first work focuses on using surrogate medical endpoint to quantify cancer mortality via automated body composition assessment; the second work utilizes deep reinforcement learning to learn better from weakly supervised dental imaging data for a finding summarization; lastly, the third work investigates federated learning settings on how they benefit cross-site collaborative learning and how non-identical and imbalanced data can impact the learning quality in real-world settings. The objective of this talk is to provide a coverage of various methods that more effectively model medical imaging tasks when the available data are constrained.
32-D463 (Star) / https://mit.zoom.us/j/5323046147
Add to Calendar
Deep Learning Tools for Next-Generation Connectomics
Location: 32-G882 (Hewlett) / https://mit.zoom.us/j/6127185653Thesis supervisor: Nir ShavitThesis committee: Polina Golland, Ed BoydenAbstract:In recent years, the field of connectomics has witnessed exciting developments. Efficient algorithms are being developed to reconstruct nanoscale maps of large-scale images, allowing us a better understanding of how neural tissue computes. However, our ability to build powerful tools for the next generation of connectomics is dependent on navigating an inherent accuracy v.s. speed v.s. scalability trade-off.This thesis addresses this tradeoff by introducing four deep learning tools and techniques applied to the acqusition, reconstruction and modeling stages of connectomics pipelines. First, we propose a way to speed up the acquisition of images using learning-guided electron microscope (EM). Second, we proposed a faster and more scalable 3D reconstruction algorithm -- cross-classification clustering (3C), for large-scale connectomics datasets. Third, we introduce a cross-modality image translation technique mapping fast X-ray images to EM images with enhanced segmentation quality. Finally, we introduced a technique to bridge the gaps between structural and functional data with connectome-constrained latent variable models (CC-LVMs) of the unobserved voltage dynamics for the whole-brain nervous system. We hope these advanced applications of deep learning techniques will help address the performance and accuracy trade-offs of next-generation connectomics studies.
32-G882 and zoom https://mit.zoom.us/j/6127185653
Add to Calendar
Zhoutong Zhang Thesis Defense: Persuing Mid-Level Vision from Casual Videos
Seminar Room G449 (Patil/Kiva)