Artificial intelligence (AI) is rapidly becoming a prominent feature in the medical device industry, with the Food and Drug Administration (FDA) having reviewed over 1,000 AI devices by the end of 2024. These devices are primarily designed to detect or triage specific health conditions. However, the industry is now shifting towards the use of broader AI tools that can analyze images, text, and various types of data across multiple contexts.
At the Radiological Society of North America’s conference, speakers focused on foundation models, which are pre-trained on massive datasets and can be adapted to various tasks. This shift towards foundation models has been noted by AI experts in the medtech and radiology sectors. Rowland Illing, Amazon Web Services’ global chief medical officer, discussed this trend and how AWS is partnering with companies like Illumina, Johnson & Johnson MedTech, Medtronic, and Abbott in the realm of AI.
Illing, who has a background in academic interventional radiology, highlighted the importance of cloud computing in implementing AI on a large scale. He emphasized the challenges of deploying AI locally across multiple medical centers with varying IT platforms, making cloud deployment a more efficient option.
The use of generative AI is on the rise, with an emphasis on building foundation models. These models, which integrate multiple data types like imaging, genomics, and language, are being developed by companies like GE Healthcare, Harrison.AI, Aidoc, and HOPPR. The integration of different data types allows for the extraction of additional information, enhancing diagnostic capabilities.
AWS is also collaborating with the FDA on leveraging generative AI to streamline the review process for drug and medical device applications. The FDA’s FiDL platform synthesizes information from companies to aid in the review process.
In the medtech field, AWS is focused on building infrastructure for foundation models. The goal is to create specialized models tailored to specific tasks, such as imaging. By utilizing large data models that consider various imaging types comprehensively, foundational models can uncover insights that may go beyond human interpretation. This shift towards foundation models represents a significant advancement in the integration of AI in the medical device industry, promising more accurate and efficient diagnostic capabilities. In collaboration with industry giants like GE, Phillips, and HOPPR, we are delving into the realm of ingesting massive amounts of data and generating insightful reports from scans. The goal is to create a foundational model for imaging that can be readily used out of the box, simplifying the process of obtaining reports from various types of scans. This model will serve as a building block for developing new applications and revolutionizing the field of medical imaging.
Once established, these foundation models can be shared with third parties, enabling them to create cutting-edge imaging applications. Whether it’s MRI, ultrasound, plain film, or CT scans, the possibilities are endless. For instance, a radiologist may spend a significant amount of time identifying abnormalities like lumps in different organs. With the help of narrow AI, specific tasks such as detecting liver masses can be automated, enhancing efficiency and accuracy.
The beauty of these foundational models lies in their extensive training on millions of images accompanied by detailed reports. This allows the models to analyze scans comprehensively, from bones to muscles to internal organs, providing a holistic assessment. While human radiologists may focus on specific areas based on clinical indications, AI can scan the entire image, offering a broader perspective.
It’s important to note that these models are not infallible and still require human oversight. Additionally, fine-tuning the model on specific datasets is crucial to ensure accuracy, especially when dealing with variations in scan protocols across different regions. By leveraging a pre-trained base model and customizing it to local datasets, we can enhance accuracy and adaptability.
Overall, the future of medical imaging lies in the seamless integration of AI-driven foundational models with human expertise. By combining the strengths of both, we can unlock new possibilities in diagnostic imaging and improve patient care.
