The Future of AI in Medical Devices: FDA Guidelines and Value of Imaging Data

Artificial Intelligence (AI) and Machine Learning (ML) are revolutionizing the medical device industry. These technologies are developing new methods to aid in the process of diagnosis, clinical decision-making, and the formation of treatment plans. AI and ML can identify anomalies and patterns to forecast patient outcomes from intricate data trends. They are making medical treatments more accurate and faster. As such algorithms evolve, they are being applied more and more in medical devices, enabling the devices to learn, adapt, and improve while being used. Medical device companies are applying these technologies to create innovative products that will better aid healthcare professionals and enhance patient care. One of the most significant advantages of AI/ML is the fact that it can learn from practical usage and experience, and improve its performance.

Evolving FDA guidance on AI in medical devices

The FDA also offers encouragement towards AI and machine learning-based software as a medical device (SaMD). In January 2021, the FDA published its "Artificial Intelligence and Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan". In this action plan, the agency describes its vision of an adaptive and iterative approach to AI regulation. The major parts of this action plan are:
‍

• Promoting Good Machine Learning Practices (GMLP) in partnership with global organizations like the International Medical Device Regulators Forum (IMDRF)
• Facilitating adaptive AI/ML systems by leveraging Predetermined Change Control Plans (PCCPs). This is a mechanism proposed for manufacturers to perform some algorithmic changes without needing to resubmit for new clearance
• Supporting algorithm transparency and promoting real-world performance monitoring post-market

On January 6, 2025, the FDA published the Draft Guidance: Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. This draft provides both lifecycle implications and particular suggestions to assist in marketing submissions for AI medical devices.

These efforts suggest that the regulatory agencies are aware of the iterative nature of AI technologies. Apart from being aware, they are also dedicated to establishing a regulatory framework that fosters innovation while promoting patient safety and trust.

Key challenges for AI and MedTech developers

The FDA's evolving guidance offers a foundation for the development of AI-enabled medical devices. But still, there are numerous challenges that remain in meeting regulatory expectations.

1. Dataset bias and lack of generalizable data

Perhaps the biggest challenge of building AI models is making sure that they are trained on data that is representative of the diversity. Models trained on non-representative data could perform suboptimal, giving rise to biased results. Addressing this concern, the FDA has emphasized making diverse datasets a part of the model-building process to minimize bias.

Furthermore, developing AI models that work well in diverse healthcare environments, populations, and imaging modalities is essential. And this will require access to large, diverse, and high-quality datasets. But these datasets are usually fragmented, proprietary, or unavailable because of privacy issues, particularly imaging data. This shortage impedes developers from training and validating models that are reliable and generalize well to actual clinical settings.

2. Privacy and data accessibility

For protection of privacy, laws like Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR) are present. These laws are deemed necessary in the healthcare industry. These regulations pose major challenges to AI and MedTech players in gaining access to real-world clinical datasets. These legislations have strict requirements around the gathering, sharing, and utilization of personal health information (PHI), which is vital in training AI models employed in medical devices.

For AI developers of medical devices, the lack of widely accessible, compliant, and annotated imaging datasets can potentially stall the progress and hamper validation efforts.

Advantages of real-world imaging data (RWiD) in AI development

Imaging data plays a fundamental role in the creation, validation, and implementation of AI-based medical devices. Diagnostics, decision support, or treatment monitoring devices alike rely on imaging datasets to demonstrate their performance, safety, and clinical significance. A few of the ways imaging data accomplishes all that:

‍

1. Improves generalizability and reduces bias

Generally, AI models are developed with limited or nuanced clinical datasets (e.g., from a limited healthcare system or demographic cohort). These models tend to perform poorly when implemented in diverse populations, devices, or clinical environments, resulting in poor consistency or diagnostic bias.

Training on varied, multi-site imaging data (scanners, demographics, care environments) assists in developing AI models that generalize more to the real-world.

For instance, a patient with pneumonia from another region might be misdiagnosed by an AI model that was trained exclusively on chest X-rays obtained from another geographic location. To improve the generalizability of its performance and satisfy regulatory requirements, the model must be trained on a versatile dataset. The dataset must consist of imaging scans from multiple global regions, diverse age groups, and assorted imaging modalities.

2. Enables regulatory-grade validation

To receive FDA clearance, AI models must demonstrate performance across diverse patient populations, settings, and imaging systems. Public datasets or single-institution data are usually insufficient for this purpose. Therefore, access to real-world imaging data linked with other sources of real-world data supports validation, showing that the model is safe and effective across geographies.

To get FDA approval for an AI tool that detects intracranial hemorrhage from brain CT scans, it is important to show data from multiple sites. This should include data obtained from different scanners, patient groups, and different presentations of hemorrhage. Real-world imaging data helps by providing access to de-identified imaging scans and linked clinical information from multiple healthcare systems, important for regulatory validation.

3. Enables deeper learning models

Developing clinically relevant AI tools for medical devices requires labeled data to train learning algorithms. Without annotations such as lesion boundaries, grading of severity, models cannot learn to identify, localize, or classify pathologies accurately. Moreover, a lack of annotated data limits the ability to demonstrate how and why the AI made a decision. This ambiguity reduces clinician trust and also causes a blockage in regulatory approval.

Annotated imaging datasets provide the basis to develop robust, interpretable, and clinically relevant AI models. These annotations, generally created by expert radiologists enable supervised training, facilitate explainability, and help meet FDA requirements for transparency and traceability.

For an AI-based device that detects lung nodules on CT scans, annotated datasets allow the model to learn not only what to detect, but also where and why. This helps create heatmaps that show the areas of concern. These heatmaps provide important insights, meeting both clinical and regulatory standards.

Experience the Segmed difference today

As regulatory frameworks continue to evolve, AI and MedTech developers need to meet requirements for how their algorithms work and the quality of data they use. The FDA’s initiatives suggest that, in the future, leveraging algorithms will be possible but only if supported by rigorous data practices and robust validation methodologies.

Real-world imaging data, especially when diverse, annotated, and longitudinal, will be central to building AI models that are safe, effective, and appropriate for regulatory validation.

Segmed, with its high-quality real-world imaging datasets, offers a unique solution that can significantly enhance the training of AI models. With this mission in mind, we provide access to 100M imaging studies from diverse modalities like X-rays, MRIs, CT scans, and ultrasounds. Our regulatory-grade, de-identified, and annotated datasets are ideal for developing AI models across oncology, neurology, and cardiology disease areas. Segmed has been part of more than 35 FDA clearances, multiple foundation models, and fit-for-purpose real-world evidence research projects. These RWiD are sourced from 2,000 healthcare locations across 5 continents. 

‍Connect with us today to explore how our diverse, tokenized imaging datasets can enhance AI model development in the medical device industry.