Design and implementation of a deep learning based CAD to empower cytology

This thesis presents a comprehensive investigation into the application of artificial intelligence and computer vision techniques to nasal cytology, with the goal of developing interpretable and clinically usable tools for diagnostic support. Nasal cytology is a valuable but underutilized method for assessing nasal mucosa conditions and inflammatory patterns. Despite its diagnostic potential, its manual nature, high inter-operator variability, and time-consuming analysis have limited its adoption in routine clinical practice. The work carried out in this thesis addresses these challenges by integrating medical expertise, machine learning methodologies, and software engineering principles into a unified, automated, and interpretable computational pipeline. The research is structured around three main contributions. The first concerns the creation of a dedicated, annotated dataset for nasal cytology, one of the first publicly available resources in this field. The dataset provides a foundation for reproducible experiments and benchmarking of modern object detection models, establishing a common reference point for future studies. Building upon this resource, the second contribution introduces a novel interpretable deep learning pipeline for automatic cell detection and classification. Leveraging prototype-based reasoning, the system not only achieves strong predictive accuracy but also o!ers visual interpretability of its decisions, enabling clinicians to understand the model’s reasoning process and validate its outputs in a transparent manner. The third contribution extends the use of artificial intelligence beyond static cytological images to the functional analysis of ciliated cells through the estimation of ciliary beat frequency (CBF). The proposed method, implemented in the DeepCilia system, combines object detection and signal processing to provide fast and accurate CBF measurements from microscopy videos. This innovation enables reproducible and objective assessment of mucociliary clearance function, with potential implications in the diagnosis of disorders such as primary ciliary dyskinesia and chronic rhinosinusitis. In addition to the methodological advances, the thesis demonstrates how these research components can be e!ectively integrated into a web-based clinical platform. The developed application provides physicians with an intuitive interface to manage patients, analyze cytological samples, visualize results, and store diagnostic data securely. The design emphasizes usability and data protection, employing server-side JWT-based authentication, HTTPS encryption, and a relational database architecture optimized for scalability and reliability. Overall, the thesis contributes to the digital transformation of nasal cytology, providing a full pipeline that spans from dataset design to deployable software. The outcomes highlight how interpretable and reliable AI systems can bridge the gap between computational research and clinical practice, paving the way for the adoption of automated cytological analysis in real-world diagnostic workflows. Future developments include multi-center dataset expansion, clinical validation, and the extension of the platform to other cytological domains. This work demonstrates that the convergence of deep learning, interpretability, and human-centered design can produce practical, trustworthy, and scalable solutions, o!ering a model for how artificial intelligence can be responsibly integrated into medical diagnostics.