Hybrid Deep Learning Architectures for the Analysis of Multi-Source Spatiotemporal Biomedical Data

点击翻译按钮获取中文摘要

Biomedical data is extremely hybrid in nature, consisting of multi-modal imaging, genomic, and non-imaging data. Parts of this data can be spatial such as a magnetic resonance imaging (MRI) scan. Other parts can be spatio-temporal such as time series of 3D volumes in MRI perfusion imaging. Genomic data can come in as a raw sequence or a summarized outcome of a genetic assay. To fully harness the potential of this data, a holistic, hybrid, and integrative approach is required. I have two long term research goals: First among them is to develop effective machine learning methodologies to enable better understanding of disparate biomedical data ultimately leading to high impact applications. My second goal is  to develop machine learning techniques to improve image acquisition devices such that they produce data which is more amenable to subsequent analysis. To fulfill this long term vision, I will pursue the following short term objectives in my Discovery program: Objective 1: Design of Hybrid Deep Learning Architectures for Perfusion Imaging and Genomic Data 1.1 Develop multi-branch recurrent and convolutional architectures for imaging and genetic data. 1.2 Develop Bayesian strategies to select architecture hyperparameters in hybrid neural nets. 1.3 Develop reinforcement learning methods to optimize neural architecture for hybrid networks. Objective 2:Design of Deep Learning Architectures for Microwave Tomographic Image Reconstruction 2.1 Develop deep neural networks to be used in conjunction with iterative 3D reconstruction methods. 2.2 Develop regularization strategies for neural networks to reconstruct directly from microwave data. 2.3 Incorporate Markov random fields (MRFs) within deep learning architectures for direct tissue segmentation from microwave data. I anticipate a high impact of my proposed research: (1) The hybrid neural architectures developed under objective 1 will be useful in a variety of biomedical applications where decisions have to be made based on multi-source, multi-format, and multi-modality data. In particular, the developed neural networks will be used for building a technology-based diagnostic tool for early detection of pseudo-progression in brain tumor patients. Pseudoprogression is a phenomenon wherein, the tumor grows in response to the treatment but then subsides. This detection is critical for treatment planning and is difficult for human experts. (2) Microwave imaging uses non-ionizing radiation unlike mammography which uses ionizing radiation, and is currently the standard screening method for breast cancer. The reconstruction algorithms designed under objective 2 would play a central role in the development of a safer breast cancer screening system.  The proposed program will also create unique opportunities to train 3 Ph.D., 2 M.Sc. and 5 undergraduate students in both theoretical and practical aspects of classical machine learning, deep learning, biomedical image analysis, and advanced computing.