Robust GPAI models for Healthcare use cases

Medical imaging foundation models have demonstrated remarkable capabilities across diverse clinical tasks, yet their deployment in real-world healthcare settings remains hampered by insufficient understanding of their robustness characteristics and failure modes. This project addresses the critical gap between the promise of general-purpose AI (GPAI) models in medical imaging and their reliable clinical implementation through a two-phase approach combining systematic robustness evaluation and targeted domain adaptation. In the first phase, we conduct comprehensive robustness analysis of existing imaging foundation models using publicly available medical imaging datasets. This analysis systematically identifies and documents model brittleness across various perturbations, distribution shifts, and edge cases commonly encountered in clinical practice. Our evaluation framework examines failure modes specific to medical imaging challenges including varying acquisition protocols, equipment heterogeneity, image quality degradation, and rare pathological presentations. The second phase leverages insights from the robustness analysis to fine-tune foundation models using curated datasets from partner hospitals across multiple clinical domains. We focus on high-impact use cases spanning surgical navigation and planning, oncological imaging assessment, histopathological analysis, and radiological interpretation. This targeted fine-tuning approach aims to enhance model resilience while maintaining broad generalization capabilities, ultimately producing robust AI systems suitable for clinical deployment. The project's outcomes will establish evidence-based guidelines for foundation model selection and adaptation in healthcare settings, provide open benchmarks for medical imaging robustness evaluation, and deliver validated models ready for prospective clinical validation across our target domains.