Fundamentals of Foundation Models in Radiological Image Processing: Opportunities, Limitations, and Clinical Case...

Artificial intelligence (AI) has become firmly established in radiology, with most current applications relying on task-specific convolutional neural networks (CNNs) such as U-Net architectures for segmentation, detection, and classification. The emergence of foundation models (FMs), however, marks a fundamental paradigm shift. These large-scale, pretrained models are designed to flexibly adapt to a wide range of radiological tasks, often requiring only minimal task-specific fine-tuning. In medical imaging, FMs are typically trained on large, heterogeneous, and multimodal data sets and enable novel applications such as zero-shot segmentation, prompt-based image exploration, and the integration of clinical context information. In parallel with academic developments, the first vendors have begun to incorporate FM-based approaches in commercial radiology products, gradually replacing narrowly specialized model architectures. For clinical practice, this technological shift offers potential benefits, including reduced false-positive findings, automated triage in emergency imaging, and improved detection of complex or atypical pathologies. At the same time, these advances come with increased demands on computational resources and costs, as well as unresolved challenges regarding validation, explainability, and regulatory approval. This review introduces the core principles of foundation models, discusses their technical and economic implications, and illustrates - using clinically relevant examples - how this new generation of models may fundamentally transform radiological image analysis.

· Unlike conventional task-specific networks such as U-Net, foundation models can be flexibly applied and adapted to different radiological tasks.. · Foundation models may improve radiological and clinical performance, particularly in complex or data-limited settings.. · Radiologists and industry partners must critically validate foundation-model outputs and ensure their explainable integration into clinical workflows in accordance with regulatory requirements..

· Shahzadi I, Borggrefe J. Fundamentals of Foundation Models in Radiological Image Processing: Opportunities, Limitations, and Clinical Case Studies. Rofo 2026; DOI 10.1055/a-2899-1290.