UCLA

The recent trend of digital pathology transition has enabled the advancement of Artificial Intelligence (AI) for complex pathology tasks. While some AI demonstrated performance comparable to human pathologists in lab studies, translating them into clinical practice remains challenging due to issues related to limitations of AI's integration into clinical decision-making, its explainability and controllability, and the reliability of AI-assisted outcomes.

To address these challenges, this thesis adopts a multi-faceted approach, combining field investigations, artifact development, and empirical validation, to study effective human-AI collaborative paradigms in digital pathology. First, it presents findings from a field study of pathologists' daily workflows, their attitudes towards AI with varying levels of automation, and recommendations for designing effective AI-assisted diagnostic systems. Second, this thesis discusses the development and validation of NaviPath, a next-generation, high-throughput AI recommendation system informed by pathologists' domain expertise, and xPath, a comprehensive and explainable AI-assisted pathology interface that seamlessly integrates with pathologists' diagnostic tasks involving multiple criteria and multimodal data. Finally, this thesis explores strategies to foster appropriate reliance on AI by harnessing pathologists' collective expertise to achieve reliable, and robust AI-assisted outcomes.

Overall, this thesis aspires to enable efficient, accurate, and safe human-AI collaborative pathology decisions -- supporting pathologists in reaching timely, cost-effective, and precise diagnoses, which can ultimately benefit patient management.