[00766] Deep learning techniques for inverse problems and imaging
Session Time & Room : 5D (Aug.25, 15:30-17:10) @G808
Type : Proposal of Minisymposium
Abstract : Inverse problems involve identifying parameters of interest from indirect data. A main challenge for solving inverse problems is that their solutions are often not well posed, i.e., not unique and/or unstable with respect to small perturbations in the data. Deep techniques have been successfully applied to a wide variety of inverse problems, especially those arising in medical imaging. The main purpose of this mini-symposium is to discuss recent developments of the deep learning techniques for solving inverse problems and the open challenges that need to be addressed in the future.
[03603] Deep learning-based medical image reconstruction from incomplete data
Format : Online Talk on Zoom
Author(s) :
Qiaoqiao Ding (Shanghai Jiao Tong University)
Abstract : Image reconstruction from down-sampled and corrupted measurements, such as fast MRI and sparse-view/low-dose CT, is mathematically ill-posed inverse problem. Deep neural network (DNN) has been becoming a prominent tool in the recent development of medical image reconstruction methods. In this talk, I will introduce our work on incorporating classical image reconstruction method and deep learning methods. The experiments on both sparse-view CT and low-dose CT problem show that the proposed method provided state-of-the-art performance.
[05360] Deep Unrolling Networks with Recurrent Momentum Acceleration
Format : Talk at Waseda University
Author(s) :
Qingping Zhou (School of Mathematics and Statistics, Central South University)
Junqi Tang (School of Mathematics, University of Birmingham)
Jinglai Li (School of Mathematics, University of Birmingham)
Abstract : Leveraging model-based iterative algorithms and deep learning, deep unrolling networks (DuNets) address inverse imaging problems. However, nonlinear problems hinder their efficacy. Our proposed recurrent momentum acceleration (RMA) framework employs a LSTM-RNN to simulate momentum acceleration, enhancing DuNets' performance. Applied to the learned proximal gradient descent (LPGD) and the learned primal-dual (LPD) methods, it results in LPGD-RMA and LPD-RMA, respectively. Experimental results on nonlinear deconvolution and electrical impedance tomography indicate significant improvements, particularly for strongly ill-posed problems.
[05368] Spatiotemoral Imaging with Diffeomorphic Optimal Transportation
Format : Talk at Waseda University
Author(s) :
Chong Chen (Academy of Mathematics and Systems Science, Chinese Academy of Sciences)
Abstract : This talk introduces a concept called diffeomorphic optimal transportation (DOT), which combines the Wasserstein distance with Benamou--Brenier formula in optimal transportation and the flow of diffeomorphisms in large deformation diffeomorphic metric mapping. Using DOT, a new variational model for joint image reconstruction and motion estimation is proposed, which is suitable for spatiotemporal imaging involving mass-preserving large diffeomorphic deformations. The performance is validated by several numerical experiments in spatiotemporal tomography.
[05387] Self-supervised deep learning for imaging
Format : Talk at Waseda University
Author(s) :
Hui Ji (National University of Singapore)
Abstract : Deep learning has proved to be a powerful tool in many domains, including inverse imaging problems. However, most existing successful deep learning solutions to these inverse problems are based on supervised learning, which requires many ground-truth images for training a deep neural network (DNN). This prerequisite on training datasets limits their applicability in data-limited domains, such as medicine and science. In this talk, we will introduce a series of works on self-supervised learning for solving inverse imaging problems. Our approach teaches a DNN to predict images from their noisy and partial measurements without seeing any related truth image, which is achieved by neuralization of Bayesian inference with DNN-based over-parametrization of images. Surprisingly, our proposed self-supervised method can compete well against supervised learning methods in many real-world imaging tasks..
