Reliable and Explainable AI for Automated Cell Counting

Computer vision and deep learning have achieved remarkable success across a wide range of applications, fundamentally transforming how visual information is processed and understood. In the domain of biomedical image analysis, the integration of these technologies is of immense importance. Machine learning and computer vision offer the potential to automate tedious manual tasks, accelerate scientific discovery, and improve clinical diagnostics by rapidly processing complex microscopy and medical imaging data. 

Despite these promising advancements, adapting the success of general-purpose computer vision to biomedical image analysis presents significant challenges. A primary hurdle is data scarcity; unlike natural images, acquiring and annotating specialized biological datasets requires extensive domain expertise, resulting in small datasets that often lack necessary morphological diversity. Furthermore, deep learning models typically operate as opaque ``black boxes.'' This lack of model interpretability is a critical barrier in healthcare and biological research, where understanding the reasoning behind an automated decision is essential for clinical verification, ensuring safety, and building trust among medical professionals.

A quintessential example of these challenges is automated cell counting, a fundamental task in regenerative medicine, pathology, and stem cell research. Accurate cell quantification is crucial for evaluating cell cultures, tracking differentiation, and monitoring disease progression. However, developing reliable automated systems is complicated by varying staining protocols, dense cellular clustering, and diverse cell morphologies. While Density Map Estimation (DME)---a regression technique that predicts object density per pixel---has emerged as a leading approach, current DME models still struggle with limited training data and inherent architectural opacity, hindering their broader deployment in sensitive biological applications.

This dissertation addresses these critical barriers by proposing novel methods to enhance both the reliability and interpretability of automated cell counting. To improve reliability and tackle data scarcity, we first introduce IDCIA, a specialized, diverse dataset of rat stem cell images featuring multiple antibody markers. This dataset establishes a comprehensive benchmark by evaluating baseline models from both the cell counting and crowd counting literature. These evaluations reveal that existing models struggle to generalize to the morphological complexities of specialized stem cell imagery. Furthermore, to address scalability and the adaptation of foundation models to microscopy, we propose the large-scale CellFMCount dataset and the SAM-Counter architecture. SAM-Counter effectively adapts the robust encoder of the Segment Anything Model for precise local density regression with minimal fine-tuning. Experimental results demonstrate that SAM-Counter achieves state-of-the-art counting accuracy while dramatically reducing the need for extensive task-specific training data.

To achieve interpretability through self-explaining models, we introduce two frameworks that provide intrinsic transparency by design. First, CountXplain integrates a prototype layer directly into the network, allowing the model to explicitly learn and base its predictions on biologically relevant visual patterns. Expert-driven validations confirm that CountXplain learns biologically sound features, although our analysis highlights an inherent tradeoff between counting accuracy and faithful interpretability. Second, to resolve issues of technical unfaithfulness caused by signed prediction cancellation in prior methods, we propose Proto4DME. This framework utilizes a non-negative prototype density decomposition, an optimal-transport coverage objective, and a novel mechanism to automatically determine the required number of prototypes. This ensures that model explanations are strictly additive, auditable, and mathematically faithful. Experiments across diverse microscopy benchmarks validate that Proto4DME successfully prevents prediction cancellation while maintaining competitive counting performance. Ultimately, the contributions of this thesis advance the development of transparent, scalable, and trustworthy artificial intelligence systems for biomedical image analysis.

Committee: Wallapak Tavanapong, Qi Li, Mengdi Huai, Robyn Lutz, and Donald Sakaguchi