New Deep Learning Methods for Medical Image Analysis and Scientific Data Generation and Compression

Medical image analysis is critical to biological studies, health research, computer- aided diagnoses, and clinical applications. Recently, deep learning (DL) techniques have achieved remarkable successes in medical image analysis applications. However, these techniques typically require large amounts of annotations to achieve satisfactory performance. Therefore, in this dissertation, we seek to address this critical problem: How can we develop efficient and effective DL algorithms for medical image analysis while reducing annotation efforts? To address this problem, we have outlined two specific aims: (A1) Utilize existing annotations effectively from advanced models; (A2) extract generic knowledge directly from unannotated images.

To achieve the aim (A1): First, we introduce a new data representation called TopoImages, which encodes the local topology of all the image pixels. TopoImages can be complemented with the original images to improve medical image analysis tasks. Second, we propose a new augmentation method, SAMAug-C, that lever- ages the Segment Anything Model (SAM) to augment raw image input and enhance medical image classification. Third, we propose two advanced DL architectures, kCBAC-Net and ConvFormer, to enhance the performance of 2D and 3D medical image segmentation. We also present a gate-regularized network training (GrNT) approach to improve multi-scale fusion in medical image segmentation. To achieve the aim (A2), we propose a novel extension of known Masked Autoencoders (MAEs) for self pre-training, i.e., models pre-trained on the same target dataset, specifically for 3D medical image segmentation.

Scientific visualization is a powerful approach for understanding and analyzing various physical or natural phenomena, such as climate change or chemical reactions. However, the cost of scientific simulations is high when factors like time, ensemble, and multivariate analyses are involved. Additionally, scientists can only afford to sparsely store the simulation outputs (e.g., scalar field data) or visual representations (e.g., streamlines) or visualization images due to limited I/O bandwidths and storage space. Therefore, in this dissertation, we seek to address this critical problem: How can we develop efficient and effective DL algorithms for scientific data generation and compression while reducing simulation and storage costs?

To tackle this problem: First, we propose a DL framework that generates un- steady vector fields data from a set of streamlines. Based on this method, domain scientists only need to store representative streamlines at simulation time and recon- struct vector fields during post-processing. Second, we design a novel DL method that translates scalar fields to vector fields. Using this approach, domain scientists only need to store scalar field data at simulation time and generate vector fields from their scalar field counterparts afterward. Third, we present a new DL approach that compresses a large collection of visualization images generated from time-varying data for communicating volume visualization results.