DESCRIPTION:

This talk will explore methods for developing trustworthy and interpretable deep learning systems for digital pathology, with specific applications to hematopathology image analysis. The first section will benchmark state-of-the-art vision transformer foundation models on hematopathology datasets. The second section will present uncertainty quantification methods that enable models to flag out-of-distribution samples and express confidence in predictions. The final section will demonstrate how biological priors can be integrated into deep learning models. These approaches collectively address the critical need for trustworthy AI systems.

LEARNING OBJECTIVES

• Assess the current capabilities and limitations of foundation models for hematopathology image analysis.
• Explain how uncertainty quantification methods enhance model transparency by detecting domain shifts and quantifying prediction confidence.
• Describe how biological priors can be integrated into deep learning architectures.