余坤興 (Kun-Hsing Yu)

Abstract:

Artificial intelligence (AI) is rapidly reshaping the landscape of cancer pathology research and clinical practice. Recent advances in high-resolution whole-slide image digitization, large-scale data curation, multi-modal machine learning algorithms, and scalable computing infrastructure have enabled the development of AI systems capable of learning directly from raw histopathology images at an unprecedented scale. These innovations allow models to capture both fine-grained cellular morphology and higher-order tissue architecture, linking visual patterns to clinical outcomes. As a result, pathology is evolving from a largely qualitative discipline into a data-rich, quantitative science, where reproducible and scalable computational methods can augment human interpretation. This transformation is opening new opportunities for biomarker discovery, risk stratification, and precision oncology, while also raising important questions about generalizability and reliability in real-world deployment.

In this talk, I will present a unified framework for next-generation pathology AI centered on three key pillars: scale, safety, and social responsibility. First, I will highlight recent progress in pathology foundation models that leverage weakly supervised and self-supervised learning to extract generalizable representations from tens of thousands of whole-slide images spanning multiple cancer types and clinical settings. By learning from heterogeneous datasets without requiring exhaustive manual annotations, these models can capture shared and disease-specific morphological features that transfer across tasks. They enable a wide range of downstream applications, including tumor detection, subtype classification, molecular characterization, and prognostic prediction, while demonstrating robustness to variations in staining, scanning, and sample preparation across institutions and populations.

Second, I will discuss advances in uncertainty-aware AI systems designed to enhance the safety and reliability of clinical decision-making. By integrating Bayesian inference, deep ensembles, and distribution-aware modeling techniques, these approaches provide calibrated estimates of prediction confidence and explicitly quantify different sources of uncertainty. Importantly, they can identify out-of-distribution cases, such as rare tumor types or diagnostically ambiguous samples, and flag them for further expert review. Such capabilities are essential for high-stakes clinical applications, where overconfident errors can lead to inappropriate treatment decisions and adverse patient outcomes.

Third, I will introduce emerging methods for fairness-aware pathology AI that explicitly address performance disparities across demographic groups and healthcare settings. Using contrastive learning and large-scale multi-cohort analyses, these approaches learn representations that are less sensitive to spurious correlations while preserving clinically relevant biological signals. They substantially reduce performance gaps across subpopulations and provide new insights into how biological heterogeneity, such as differences in somatic mutation prevalence, and data imbalance jointly contribute to variability in AI performance.

Finally, I will outline key challenges and future directions in building robust, generalizable, and clinically deployable AI systems, including issues of transportability across populations, interpretability of complex models, rigorous external validation, and seamless integration into clinical workflows. Together, these ongoing advances point toward a new paradigm in pathology, where AI not only augments diagnosis but also drives scientific discovery, democratizes access to high-quality clinical care, and supports safer and more informed decision-making.