Computational pathology in breast cancer: optimizing molecular prediction through task-oriented AI models

Area of Science:

• Pathology
• Artificial Intelligence
• Oncology

Background:

• Foundation models in AI for breast cancer pathology analyze whole slide images (WSIs) for diagnostic and prognostic insights.
• Clinical adoption of these large AI models is hindered by poor workflow integration, limited explainability, and lack of generalizability.

Purpose of the Study:

• To examine the potential of small, task-oriented AI models for predicting key molecular features in breast cancer directly from WSIs.
• To address the limitations of foundation models by exploring techniques like model distillation and weak supervision.

Main Methods:

• Focus on task-specific AI models for predicting hormone receptors (HRs), HER2, Ki-67, BRCA-related status, and somatic mutations from WSIs.
• Critical evaluation of methods such as model distillation, weak supervision, and modular training to enhance AI model performance and applicability.

Main Results:

• Small, task-oriented AI models show promise in predicting clinically relevant molecular features directly from breast cancer WSIs.
• Techniques like model distillation and weak supervision can help overcome the constraints of large foundation models.

Conclusions:

• Task-oriented AI models offer a viable path toward integrating AI into breast cancer pathology, providing explainable and clinically actionable insights.
• Advancement requires high-quality datasets, multi-institutional validation, and interdisciplinary collaboration among computational scientists, clinicians, and regulators.