Revisiting mitosis quantification in cancer in the era of AI

Cancer is characterized by uncontrolled cell proliferation, often driven by aberrant mitotic activity. Mitosis, a fundamental process in the cell cycle, is frequently dysregulated in cancer due to genetic and environmental factors, leading to unchecked growth and tumor progression. In histopathology, mitotic figure counting plays a pivotal role in tumor grading and prognostication. However, manual mitosis counting remains labor-intensive, prone to inter-observer variability, and challenging due to the difficulty of distinguishing mitotic figures from morphological mimickers. Recent advances in computational pathology have introduced automated mitosis detection methods using deep learning (DL), offering faster, more consistent, and scalable solutions. Nevertheless, domain shift (DS) across histopathology datasets, stemming from variations in staining protocols, scanner technologies, and tumor types, presents a major hurdle in the generalization of DL models. This thesis aims to develop a domain-generalizable and efficient mitosis detection framework that can address DS challenges and facilitate large-scale downstream analyses of mitotic activity in cancer. First, we systematically investigate DS in computational pathology (CPath) and benchmark various domain generalization (DG) techniques. By evaluating 28 cutting-edge DG algorithms, we provide a comprehensive framework for mitigating DS and improving model robustness in histopathology applications. Next, we propose a two-stage mitosis detection framework, MDFS (Mitosis Detection Fast and Slow), which combines a segmentation-based detection stage with a high-precision classification refinement step. This method achieves state-of-the-art performance on multiple public datasets and secured first-place rankings in the MIDOG 2021 and 2022 challenges. Additionally, we evaluate the necessity of self-supervised pretraining and various augmentation techniques, optimizing the model for real-world clinical deployment. Leveraging MDFS, we perform large-scale WSI-level mitosis quantification across 9,141 whole-slide images (WSIs) from 31 cancer types in The Cancer Genome Atlas (TCGA). Unlike conventional hotspot-based mitosis counting, we introduce a novel mitotic network analysis approach based on social network analysis (SNA) techniques. This enables the extraction of mitotic topological features (MTFs) that characterize mitotic spatial distribution, connectivity, and heterogeneity within the tumor microenvironment (TME). Our findings demonstrate that mitotic density, clustering, and spatial heterogeneity are strong prognostic biomarkers independent of classical clinical factors. Furthermore, we introduce an atypical mitosis detection model to quantify mitotic errors, revealing their significant association with genomic instability, proliferation pathways, and immune modulation across multiple cancer types. This thesis provides a robust framework for mitosis detection and analysis, addressing critical challenges in DS while expanding the role of mitotic activity as a prognostic and biological marker in cancer research. By integrating DL-based mitotic profiling with genomic and clinical data, this work paves the way for future precision oncology applications. Our publicly available dataset of over 5 million mitotic figures, enriched with mitotic topology metrics, serves as a valuable resource for advancing research in computational pathology and cancer biology. The methodologies and insights presented here mark a significant step towards the broader adoption of AI-driven solutions for automated histopathological assessment, ultimately improving cancer diagnostics and patient outcomes.