AI Foundation Models Enhance Diagnostic Insights in Oncology

The diagnostic process in oncology frequently involves the integration of diverse data streams, including imaging, pathology, genomics, and clinical history. The sheer volume and complexity of this information can lead to diagnostic delays or inaccuracies, particularly when presentations are atypical or when rare mutations are involved. Traditional diagnostic algorithms often struggle with the combinatorial explosion of potential insights derived from multi-modal data. The discussions at ASCO 2026 centered on how AI foundation models, trained on vast, heterogeneous datasets, are being developed to address these limitations by identifying subtle patterns and generating diagnostic hypotheses that may elude human interpretation or conventional rule-based systems.¹

These models are distinct from earlier AI applications in oncology, which often focused on specific tasks, such as image segmentation or mutation detection. Foundation models, by contrast, are designed for broad applicability, capable of performing multiple tasks and adapting to new data with minimal retraining. Their architecture allows for the identification of complex relationships across different data modalities, potentially leading to more comprehensive diagnostic profiles. For instance, a model might correlate specific histopathological features with genomic alterations and patient demographics to predict treatment response or disease progression with greater precision than any single data type could provide alone.¹

What the presentations highlighted

While specific trial data were not detailed in the general overview, the ASCO 2026 presentations emphasized the conceptual shift these models represent. The focus was on their ability to act as 'diagnostic assistants,' providing clinicians with a more holistic view of a patient's disease. This includes the potential to identify patients at higher risk of recurrence, predict resistance to standard therapies, or suggest novel therapeutic targets based on an integrated understanding of their unique biological profile. The models are posited to improve the identification of rare cancers and uncommon subtypes, where diagnostic expertise may be geographically limited or where the diagnostic journey is often protracted.²

The utility of these foundation models extends beyond initial diagnosis to ongoing patient management. By continuously integrating new clinical data, they could theoretically refine prognostic assessments and adapt treatment recommendations in real-time. This iterative process aims to reduce the incidence of misdiagnosis or delayed diagnosis, which are known contributors to suboptimal patient outcomes and increased healthcare costs. The discussions also touched upon the ethical considerations and regulatory pathways required for the clinical implementation of such advanced AI systems, acknowledging the need for rigorous validation in prospective, real-world settings.²

Limitations of current AI foundation models include their 'black box' nature, where the reasoning behind a diagnostic recommendation may not be transparent, posing challenges for clinical accountability and trust. The generalisability of models trained on specific populations to diverse global cohorts also requires careful evaluation to prevent algorithmic bias. Furthermore, the computational resources required to train and deploy these models are substantial, which could limit their accessibility in resource-constrained environments. Future work will need to focus on developing interpretable AI, ensuring data privacy and security, and establishing robust regulatory frameworks that can keep pace with rapid technological advancements while safeguarding patient safety.³