Foundation models in clinical oncology: Progresses and perspectives

Area of Science:

• Oncology
• Artificial Intelligence
• Biomedical Data Science

Background:

• Cancer presents a significant global health challenge, with rapidly increasing volumes of diverse data (pathology, radiology, genomics, EHRs).
• Foundation models (FMs) are emerging as powerful tools for cancer screening, diagnosis, outcome prediction, and treatment selection.

Purpose of the Study:

• To review the potential of foundation models in oncology.
• To identify key barriers hindering the development and implementation of FM-enabled cancer care tools.
• To propose strategies for overcoming these challenges.

Main Methods:

• Review of current applications and challenges of foundation models in oncology.
• Analysis of data heterogeneity, annotation scarcity, regulatory hurdles, privacy issues, and potential biases.
• Discussion of collaborative approaches and data infrastructure requirements.

Main Results:

• Foundation models offer transformative potential across various oncology tasks, from diagnosis to clinical trial design.
• Significant barriers include data variability, limited high-quality annotated cohorts, evolving regulations, data privacy, and algorithmic bias.
• Cross-institutional partnerships and robust data infrastructure are essential for progress.

Conclusions:

• Addressing data quality, standardization, privacy, and equity is paramount for realizing the full potential of FMs in cancer care.
• Collaborative efforts embedding AI expertise within clinical teams are vital for developing generalizable and trustworthy AI solutions.
• Strategic integration of FMs promises to reshape the future of oncology practice.