Scalable Learning for High Dimensional Imaging Data

Abstract

The exponential growth of high-dimensional imaging data presents major opportunities and computational challenges across scientific and biomedical domains. This dissertation addresses scalable learning for three imaging modalities: resting-state functional magnetic resonance imaging (rs-fMRI), multispectral satellite imagery in remote sensing archaeology, and whole-slide kidney pathology images. For rs-fMRI, we analyze “avalanches,” or peaks in the global signal, and show that these events may encode meaningful physiological information rather than noise. For remote sensing archaeology, we introduce DeepAndes, a vision foundation model pretrained with DINOv2 self-supervised learning on three million multispectral satellite image patches from the central Andean region. DeepAndes shows clear scaling behavior and outperforms supervised and self-supervised baselines in loci classification, image retrieval, and few-shot segmentation. For renal pathology, we develop a human-in-the-loop framework that integrates multiple cell nuclei foundation models, including Cellpose, StarDist, and CellViT, to scale up high-quality annotations and fine-tune kidney-specific segmentation models. Collectively, these contributions advance scalable and domain-adaptive learning for high-dimensional imaging.