Towards spatially-lucid pattern discovery

Abstract

Spatial data science has transformed modern life through innovations such as navigation systems (e.g., Google Maps, eco-routing), precision agriculture (e.g., GEOGLAM), land use monitoring, and autonomous driving technologies (e.g., high-definition maps). These advances have reshaped industries and addressed global challenges. These advances have reshaped industries and opened up trillions in potential annual revenue by the end of the decade, with agriculture and transportation among the largest contributors. Recently, spatial data science has expanded to the cellular and molecular level through technologies such as multiplex immunofluorescence (MxIF) imaging. Researchers are now able to map the locations of immune and tumor cells within tissue biopsies, as demonstrated by efforts like the Human Cell Atlas. These developments illustrate the growing impact of spatial data science on scientific discovery, economic opportunity, and societal benefit. Spatial data properties pose significant challenges for traditional data science and artificial intelligence (AI) techniques. The first law of geography, spatial auto-correlation, states that nearby entities tend to be more closely related than distant ones. This inherent dependency requires models to account for complex spatial relationships rather than relying on simple distance measures. For example, it needs to account for spatial arrangements of nearby objects. Additionally, spatial autocorrelation varies in strength and direction, which calls for methods that can prioritize significant spatial arrangements. The second law, spatial heterogeneity, highlights differences in spatial properties across locations due to environmental, social, or structural variation. These properties violate the foundational assumption of many machine learning algorithms that the data is identically and independently distributed. Additional challenges include limited labeled data, especially in healthcare, where annotation requires domain expertise, and the need for spatially-explainable models for high-stakes use cases. My thesis addresses these challenges by developing novel spatially-lucid AI classification methods that can accommodate the unique characteristics of spatial data. These methods are designed to distinguish between two groups based on spatial concepts, relationships, and patterns. For instance, oncologists are highly interested in identifying spatial co-location patterns, that is, subsets of cell types that frequently occur nearby, as these patterns may help explain differences in clinical outcomes, such as distinguishing responders from non-responders to a candidate therapy. My research focuses on analyzing biopsy data, specifically cellular maps derived from multiplex immunofluorescence (MxIF) imagery, to investigate the spatial underpinnings of these clinical responses. First, this thesis proposes a spatial-interaction-aware multi-category deep neural network (SAMCNet) for spatial-configuration-aware AI classification. The model captures significant arrangements among nearby spatial objects (e.g., cells) with varying interaction strengths and incorporates two key components: local reference frame characterization and point-pair prioritization layers. SAMCNet is an example of a spatial influencer model that uses distinguishing location-dependent spatial arrangements as input and produces accurate location-dependent classification as output, capturing spatial variation in the data. Second, this thesis refines the spatial-configuration-aware AI classifier by allowing network parameters to vary across different place types (e.g., tumor regions such as the tumor core and the interface) rather than relying on traditional one-size-fits-all scalar approaches. It introduces a place-calibrated AI classification framework that incorporates concepts such as k-nearest neighbors, place adaptation, a weighted-distance learning rate, place-independent layers, and place-dependent layers to capture variation in inputs, outputs, and model parameters. Finally, this thesis proposes a method that incorporates distinguishing regional spatial arrangements (i.e., co-location patterns) into an unsupervised domain adaptation classifier within a multi-task architecture. The model emphasizes spatial arrangements and utilizes spatially oriented pretext tasks to enhance domain adaptation performance on multi-type point maps, such as cellular maps. Extensive experimental results on real-world datasets (e.g., oncology data) show that the proposed approaches provide higher prediction accuracy than baseline methods. Real-world case studies demonstrate that the proposed approaches discover patterns that are missed by the existing methods and have the potential to inspire new scientific discoveries.