Methods for Automatic Processing and Visual Annotation of Microscopy Images

Abstract

This thesis advances automated analysis of biomedical microscopy images through an integrated computational ecosystem addressing critical challenges in Transmission Electron Microscopy (TEM) of immune cells and Whole Slide Imaging (WSI) for histopathology. We introduce the TEM-NeutroStruct dataset, the first comprehensive resource for neutrophil intracellular structures, comprising 93 high-resolution TEM images with 14,918 expert-validated annotations across seven classes with novel quantitative classification criteria based on size and intensity statistics, transforming traditionally subjective visual assessments into reproducible, evidence-based standards. To accelerate dataset expansion while maintaining biological accuracy, we develop a semi-automatic annotation framework integrating YOLO object detection with CVAT in a human-in-the-loop workflow, reducing manual annotation workload by 80% through iterative expert refinement and U-Net-based central cell isolation (validation loss 0.050). For robust automated detection, we introduce Dynamic Edge Annotation Assessment (DEAA), a novel preprocessing method that dynamically evaluates edge annotations based on proximity, aspect ratio, and area; combined with multi-scale YOLOv9 training, DEAA boosts detection performance from mAP@50 of 0.688 to 0.859, achieving 87.62% overall counting accuracy with 98.95% accuracy for Primary Granules. To ensure trustworthy automation, we develop Morphology-Regularized Hierarchical Conformal Prediction (MR-HCP), the first uncertainty quantification framework combining hierarchical biological taxonomies with morphology-aware nonconformity scores, providing rigorous coverage guarantees (95.4% at 90% target) with compact prediction sets (85.2% singletons, 93.4% singleton accuracy) enabling efficient expert-in-the-loop workflows. For cross-modality applicability, we present HistoMSC for WSI histopathology analysis employing Morse-Smale Complex topology and Density Comparison Graphs, demonstrating multi-scale visualization that reduces cognitive load while automating tissue segmentation. These contributions establish a complete analytical pipeline—from dataset creation with objective criteria through efficient annotation and robust detection to rigorous uncertainty quantification—advancing neuroimmune research and providing scalable tools for cellular ultrastructure analysis in immune cell biology.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering