Domain-aware and parameter-efficient computer vision for multi-domain scientific image analysis

Abstract

Scientific image analysis plays a critical role in understanding complex phenomena across multiple disciplines, like environmental monitoring, structural biology, and medical imaging, yet applying modern computer vision to these domains remains fundamentally challenging. Unlike natural images, scientific images are characterized by domain-specific noise, extremely low signal-to-noise ratios, subtle physically determined structural cues, and annotation processes requiring specialized expert knowledge that is scarce and expensive to obtain at scale. These characteristics create three closely related challenges: limited annotated data, differences across imaging domains, and complex visual structures. Standard deep learning models, and even large vision foundation models, often struggle to handle these challenges effectively when applied directly to scientific imaging tasks. This dissertation investigates how domain-aware computer vision frameworks, grounded in the combination of pretrained visual representations, domain-specific structural inductive biases, and parameterefficient adaptation, can be designed to overcome these shared challenges across diverse scientific imaging contexts. First, this dissertation presents a deep learning framework for estimating oil spill concentration from fluorescence images. The framework uses MobileNetV3-based feature extraction, a regression model, and uncertainty quantification to support rapid, field-oriented concentration assessment through mobile and web-based deployment, offering a practical and scalable alternative to traditional laboratory-assessment methods. Second, it introduces CryoFSL, a few-shot learning framework for automated protein particle picking in cryo-electron microscopy micrographs. CryoFSL adapts the Segment Anything Model 2 (SAM2) through lightweight parameter-efficient adapter modules inserted into a frozen hierarchical encoder, enabling reliable particle detection from as few as five annotated micrographs. This approach reduces the need for extensive particle annotations and improves the quality of particles selected for downstream three-dimensional reconstruction. Third, it proposes a SAM2-based nuclei segmentation framework with a Gated Dual-Branch Adapter (GDBA) that integrates semantic features from the frozen foundation model with frequency-domain structural features to improve boundary delineation and cross-dataset robustness in histopathological images. Together, these contributions demonstrate that scientific image analysis benefits from more than direct transfer of general-purpose models. Robust performance requires adaptation strategies that reflect the structural and statistical properties of the target domain. Across all three projects, the dissertation shows that parameter-efficient learning, structural feature integration, uncertainty awareness, and foundation-model adaptation can improve the reliability and practical value of computer vision systems under limited data and domain shift. Overall, this work advances a unified framework for domain-aware scientific image analysis and provides practical methods for environmental monitoring, structural biology, and computational pathology.