PhD Thesis Defense

Zoom link: https://caltech.zoom.us/j/82340572016

Accurate segmentation of biological structures is essential for quantitative microscopy. It allows downstream analysis of cell morphology, spatial genomics, tissue organization, and disease-associated microenvironments. However, segmentation across microscopic imaging modalities remains difficult due to limited annotated data, dense and small objects, diverse imaging platforms, heterogeneous staining, and extremely large whole-slide images. Existing AI-driven computer vision methods provide powerful foundations, but they require adaptation before reliably supporting biological image analysis.


This thesis tackles these challenges through four connected contributions. First, it establishes a universal preprocessing and evaluation framework. It bridges the gap between the biological microscopy format and the industry-standard COCO format. This framework enables dataset conversion, tiling, normalization, annotation filtering, and model benchmarking. It also fixes the problem of inconsistent evaluation caused by polygon transcoding errors for small objects. Second, this thesis investigates an equivariant segmentation architecture based on the roto-translation equivariant CNN and attention modules. It demonstrates that equivariance can improve performance under limited annotation and weak augmentation settings. Third, it contributes to CellSAM, a foundation-model-based cell segmentation framework that adapts SAM to cellular images through automatic cell detection, SAM fine-tuning, and large-scale cross-modality evaluation. CellSAM achieves strong generalization across tissue, cell culture, H&E, bacteria, yeast, and nuclear datasets. Fourth, this work extends segmentation from individual cells to functional tissue units (FTUs). FTU segmentation introduces additional challenges, including color variations across imaging platforms, variable object size, and whole-slide inference. To address these issues, this thesis proposes a multi-stage FTU segmentation design incorporating stain normalization, FTU type detection, and dynamic scaling based on object size.


These contributions collectively bridge computer vision and microscopy analysis, providing effective methodologies and model designs for robust segmentation from individual cells to FTU-scale biological structures.