Reliable and Explainable AI for Automated Cell Counting

Automated cell counting supports large scale biological analysis, yet performance is sensitive to modality shifts, cellular overlap, and limited interpretability. We begin with resources for controlled and reproducible evaluation. Our first fluorescence microscopy dataset examines how staining variability and background signal affect density map estimators across conditions. Our second fluorescence microscopy benchmark provides diverse markers, cell types, and densities with standardized splits and metrics that reveal strengths and failure modes of existing methods. Within this benchmark, we also demonstrate the adaptation of vision foundation models to automated cell counting and analyze where pretraining helps and where domain gaps remain.

Building on this foundation, we introduce a prototype guided density map framework that makes what the model counts transparent through learned cell and background prototypes. Alignment and diversity objectives encourage meaningful and non redundant concepts, which support explanations while maintaining competitive accuracy. Our ongoing work focuses on two areas. We are building models for robust counting, focusing on generalization across acquisition differences. Second, we are developing interpretability methods that automatically determine the number of prototypes and improve prototype learning to yield clearer, more faithful concepts. In sum, these studies create a comprehensive pathway to reliable and transparent automated cell counting in real laboratory practice.

Committee: Wallapak Tavanapong (major professor), Qi Li, Mengdi Huai, Robyn Lutz, and Donald Sakaguchi