The Image Processing Seminar Series is a collaboration of faculty and students conducting research related to image processing in the School of Engineering and Applied Science and across the University of Virginia. This series provides a venue for presenters to share their love of image processing with an audience of like-minded researchers.
This seminar features three talks by the Virginia Imaging Theory and Algorithms Laboratory.
- “T1-mapping of the heart with convolutional neural networks” by Haris Jeelani
Abstract: ‘T1’ is the time taken by the hydrogen atoms in the heart wall to recover to their longitudinal equilibrium state after excitation. It can be used for early detection variety of pathological conditions such as heart failure. Traditionally, a pixel-wise nonlinear model fitting that is sensitive to noise is used to obtain T1 maps of the heart. In this talk, I will discuss our approaches to increase the noise robustness using our convolutional neural network framework (DeepT1). The DeepT1 framework includes a recurrent and a U-net model. I will also discuss the approaches being taken to improve the existing DeepT1 framework. This is joint work with Dr. Michael Salerno and Dr. Christopher Kramer (Cardiology) and Dr. Yang Yang (now, Mount Sinai School of Medicine).
- “Fast and Robust Hyperparameter Tuning for Image Processing” by Tanjin Toma
Abstract: In image and video processing, algorithms for inverse problems (e.g., image enhancement, medical image reconstruction, etc.) often have some processing hyperparameters that need to be set effectively to yield good quality solutions. Existing automatic hyperparameter tuning methods are mostly iterative, thus computationally heavy and often rely on predetermined image features (such as, image quality measure, risk estimate) to estimate hyperparameter values. In this talk, we discuss our data-driven approach for hyperparameter tuning that exploits a convolutional regression network to estimate hyperparameters in one-shot once pre-trained offline.
- “Optimizing Microscope Image Analysis for Modern Neuroscience” by Tyler Spears
Abstract: Neuroscientists today require orders of magnitude more data than before, to answer questions about neurologically-based diseases, developments, and behaviors. But, an increase in data requires more automation to extract and make sense of that information. In this proposal talk, we will discuss plans for developing image processing and machine learning tools in collaboration with Dr. Jaideep Kapur (UVA Neuroscience) and Dr. Cedric Williams (UVA Psychology). These tools will form 3D representations of microscopy imaging data and locate and describe targeted neurons present in those images. We will end with a discussion on the potential impacts of accelerating these neuroscience experiments, such as in the study of learning and memory, and of neurological disorders that affect learning.