Discover how Segmed's tokenized Real-World Imaging Data (RWiD) enables the tuning of healthcare foundation model for maximizing healthcare outcomes.

Introduction to healthcare foundation models

The rapid advancement in highly adaptable and reusable foundation models promises to bring new capabilities in the field of medicine. Healthcare foundation models (HFMs) are large, pre-trained AI models used for multiple applications in healthcare. The models are first trained on massive, generalized sets of data so that they learn and capture patterns across a wide range of therapeutic areas. Healthcare foundation models provide solutions to complex tasks such as streamlining clinical workflows through information extraction from unstructured data, identifying pathologies, and generating clinical reports.

Potential applications of HFMs include real-time patient interaction through virtual medical chatbots that help in support, triage, and education. One of the major applications is improved clinical decision support by assisting with diagnoses, treatment recommendations, particularly in radiology and pathology. Also, HFMs have capabilities to perform data extraction that summarizes large volumes of real world data for more rapid insight.

Harnessing the computational ability of these models can result in a tremendous amount of improvement in diagnostic effectiveness, decision-making and optimizing patient treatment. But for these models to work effectively in a real-world setting, fine-tuning them is essential. This is especially necessary for medical imaging, which is an important source of real-world data.

‍

Significance of RWiD in fine-tuning HFM

‍Although pre-trained HFMs provide a strong foundation, their performance can be greatly improved in the real-world setting by fine-tuning them with real-world data. Training HFM models refers to the initial process of teaching a model to understand general patterns and structures from a large-scale diverse dataset. During this process a model learns by adjusting its parameters and minimizing errors. Once the model is trained, it can be used for general tasks but may not be optimized for specific applications.
‍
This is where the process of fine-tuning comes in. Fine tuning an HFM is a subsequent step done to make the model more capable of a specific task. In the fine-tuning process, a model is further trained on a smaller, task-specific dataset enhancing its accuracy and performance in that specific domain. Fine-tuning requires leveraging more specific and granular data so that it adapts to the nuances of the advanced real-world data captured in everyday healthcare practices.

In the absence of fine-tuning of HFMs, there arises a challenge in identifying and managing variation, especially when interpreting real-world imaging data.  These variations can arise due to variations in imaging modalities used, patient demographics, and the clinical context in which the image was acquired. Fine-tuning HFM using novel data sources like real-world imaging data (RWiD) enables the model to become more accurate and credible in predicting outcomes, reducing diagnostic error. This increases the effectiveness and ease of adopting HFMs in clinical practice.

‍

Benefits of using RWiD for fine-tuning HFM 

‍

1. Enhanced diagnostic accuracy 

One of the greatest advantages of integrating RWiD into HFM is enhanced diagnostic accuracy. HFMs trained using traditional sources of data (such as claims and electronic health records) tend to be narrow in scope, focusing on a limited number of variables with few sources of data. By introducing RWiD sources from diverse clinical environments, the model is exposed to a greater depth and granularity of data. This helps the models predict imaging variability, picking up on subtle, clinically relevant features.

‍For instance, the detection of lung cancer can be significantly improved through fine-tuning with RWiD. A generic data-trained foundation model may not be able to detect tumors in patients with comorbidities, where the CT scans may have overlapping characteristics. But when the model is fine-tuned using RWiD (actual case CT scans), it becomes more capable of detecting tumors even with such confounding factors.

‍

2. Improved model robustness and generalization

Including RWiD in the fine-tuning procedure improves not just model accuracy but also the robustness and generalizability of healthcare foundation models. The model is exposed to diverse and varied clinical scenarios such as patient demographics and disease presentation, helping the model to adapt to variations in data. Further, RWiD enhances model generalization by ensuring the model can perform across different healthcare settings and recognize subtle patterns across diverse patient groups.

‍‍

3. Improved precision medicine and predictive analysis 

Refining HFM with RWiD is central to improving personalized medicine. By integrating diverse patient-specific imaging data into HFM, the model is able to assist in providing treatments that address individual variability. Moreover, RWiD-tuned models improve predictive analytics to allow healthcare providers to predict health trends and inform preventive measures, leading to more effective and targeted interventions.

‍

Segmed’s role 

‍Segmed, with its high-quality real-world imaging datasets, offers a unique solution that can significantly enhance the training of Healthcare Foundation Models (HFMs). We offer access to 100 million de-identified imaging studies from diverse modalities like X-ray, MRI, CT, and ultrasound. These RWiD are sourced from 2000 healthcare locations across 5 continents. 

By employing Segmed's RWiD solutions, AI engineers can tune models to adapt to the complexity and intricacies of real-world imaging data. This encompasses providing solutions for major use cases like improved diagnostic accuracy, enhanced generalizability across populations, and improved robustness to varying clinical settings.

Connect with us to explore how our diverse, high-quality tokenized imaging datasets can enhance the training and validation of healthcare foundation models. Contact us to discover how RWiD can enhance the efficiency and optimization of healthcare foundation models.



Sources

Lin C, Kuo CF. Roles and Potential of Large Language Models in Healthcare: A Comprehensive Review. Biomed J. 2025 Apr 29:100868.
‍
Mahesh N, Devishamani CS, Raghu K, Mahalingam M, Bysani P, Chakravarthy AV, Raman R. Advancing healthcare: the role and impact of AI and foundation models. Am J Transl Res. 2024 Jun 15;16(6):2166-2179.
‍
Khan MO, Afzal MM, Mirza S, Fang Y. How Fair are Medical Imaging Foundation Models? Proc of Machine Learning Research. 2023; 225:1-13.
‍
Bassuk I, Lin T. How Did We Get Here? The Evolution of Foundation Models in Healthcare.
‍
Merkow J. Discovering the Power of Finetuning MedImageInsight on Your Data. Microsoft Tech Community. 2025 Mar 24.