---

As generative artificial intelligence (genAI) continues progressing into mainstream research pipelines, the opportunities available for its application to supporting clinical decision research do as well. Here we aim to provide a brief perspective on how genAI methods and mechanistic models may be integrated into a mechanistic learning framework. While recently most attention has been directed towards genAI approaches applied to natural language, in particular large language models, in this post we focus on the generative application to medical images.

Towards this goal, we provide a perspective on how one modeling paradigm (genAI or mechanistic models) can complement the other, highlighting the opportunities created by this integration. Within the framework of continuous advancements in modeling approaches and genAI in particular, we reflect on their influence on the role of a mathematical oncologist in the clinical team.

Mechanisms to Machine Learning:

Machine learning, and deep learning in particular, rely on large quantities of high-quality data to be trained effectively for a task. By using mechanistic modeling approaches, an enhancement in the machine learning results can be achieved by enforcing domain-specific knowledge and enhancing data-driven models. By allowing for more structured and defined priors there is potential for an improved interpretability and generalizable models.

Not only do mechanistic models contribute through limiting the search space of machine learning models, but they can also provide complementary information. This is enabled by the intrinsically different data requirements. Let us consider an imaging-based genAI model aiming to generate a brain magnetic resonance imaging (MRI) scan and predict how a tumor will appear at a later time point (Fig. 1). A single high-quality complete imaging session may be enough to abstract the patient-specific anatomy, however insufficient to approximate tumor dynamics. In such limited data scenarios, either due to their quality or quantity, the approximation of future tumor burden using imaging data alone is not possible. However, mechanistic models have the advantage of modeling tumor burden only relying on available volumetric measurements, accessible sometimes even in low quality data scenarios. Taken together, an integration of the two approaches could prove essential for advancing current genAI (and not only), by embedding mechanistic insights into the learning process.

Machine Learning to Mechanisms:

Similarly, the inverse approach is possible by identifying hidden patterns within highly complex, high-dimensional data. Through the analysis of large datasets, the potential of artificial intelligence includes the suggestion of missing dynamics, parameter estimation and functional form. The revealing of the additional information through analysis of high-dimensional parameter spaces could help refine existing models or uncover new governing equations. Practically, automated segmentation, volumetric measurement, and radiomic analysis of tumors from medical images have now been significantly accelerated and enhanced by machine learning algorithms. With advances in imaging techniques such as magnetic resonance imaging, in many applications, including adult and pediatric high-grade brain tumors[1], these precision imaging analyses can now reliably provide critical, quantitative inputs such as parameters required to calibrate mechanistic models of tumor growth and progression. Moreover, genAI has unique capability in addressing data scarcity by synthesizing missing intermediate time-point images (Fig. 2), thereby enriching longitudinal datasets necessary for mechanistic modeling.

Aiding the Mathematical Oncologist:

The role of the mathematical oncologist can be significantly enhanced by the application of genAI into the current workflow. Mechanistic learning can improve predictive modeling, as well as accelerate the development of personalized treatment strategies - allowing for actionable decisions to be made within the multidisciplinary care team. This can further apply to the application of optimal clinical trial design, allowing for the prediction of subpopulations most likely to respond favorably to treatment. Through these enhanced workflows, the translation of complex outputs into interpretable clinical recommendations becomes ever more streamlined, strengthening the bridge between current modeling approaches and patient care.

Overall, integrating genAI into current mathematical-oncology workflows may enhance both discovery rates and validation, improving applications within clinics. Still, as with all newer technologies – maintaining healthy skepticism and critical perspective is essential; outputs of these enhanced approaches must undergo rigorous review but both researchers and clinical staff prior to any implementation. We look forward to watching future work and direction of this exciting approach.

References

1. Z. Jiang et al. Magnetic Resonance Imaging Feature-Based Subtyping and Model Ensemble for Enhanced Brain Tumor Segmentation. (2024). doi: 10.48550/arXiv.2412.04094