AI/ML Articles & Videos

This is a recording of the talk Yujan gave at the RSNA AI Theater

Ideas on how to get foundation models and generative AI FDA cleared.

The discussion explores generative AI (GenAI) in medical devices, focusing on FDA's considerations for regulation, risks, and the lifecycle management of such technologies. Key points include the complexity of training data in foundation models, the risks of hallucinations in outputs, and the evolving role of post-market surveillance. Insights highlight challenges in balancing innovation with regulatory oversight and safety.

Dr. Elena Sizikova's presentation highlights the role of synthetic data in addressing data scarcity and privacy challenges in medical imaging AI. Synthetic data offers benefits like customizable variability, reduced risks, and improved bias control, particularly in fields like histopathology and pediatric imaging. She contrasts generative and knowledge-based models, emphasizing the potential of hybrid datasets for enhancing AI performance. The talk also stresses the importance of robust evaluation frameworks to validate synthetic data's effectiveness.

Yujan and David discussed the topic of FDA's performance in regulating AI medical devices, focusing on the ideal regulatory situation, the impact of AI on risk analysis, and the regulatory pathways for adding AI tools to existing software medical devices. They also explored the potential of continuous learning models in medical devices, the adoption of AI and ML in the US healthcare market, and the cybersecurity threats facing medical devices. The conversation ended with a discussion on the regulatory implications of continuous learning algorithms and the potential for a letter to file for a SAMD cleared by FDA for surgical planning.

A 10x talk exploring the basics of prompt engineering, followed by an open-ended discussion.

Recent advancements in radiology have led to the development of foundation models integrating vision and language capabilities, transforming medical imaging analysis

The talk covers device changes and their regulatory implications using FDA guidance. It focuses on two key documents for deciding when a 510(k) submission is needed for hardware and software changes. A "letter to file" process is used for post-market evaluations, while the guidance also helps in pre-market strategies for anticipating design changes that may require a 510(k).

The article outlines a method for documenting AI/ML algorithms for FDA pre-submissions. It emphasizes clear algorithm descriptions to avoid delays. Key steps include visual runtime descriptions, dataset analysis, non-ML testing, annotation details, AI/ML test plans, and performance metrics. It stresses reducing annotation costs and leveraging existing data while advising comprehensive documentation for effective FDA evaluation.

With predetermined change control plans (PCCPs), FDA has given manufacturers a great new regulatory tool. In this article, we discuss PCCP best practices and questions based on our experiences with FDA and our thorough research of the FDA databases.

Tips and FAQs on how to get a foundation model FDA 510(k) or De Novo cleared.

What is the difference between user needs and requirements?

Learn how to avoid invalidating your test dataset while still allowing yourself some flexibility to re-use the test data.

This video discusses the potential avenues for getting generative AI applications cleared through FDA.

This article outlines the process of developing an AI/ML algorithm from scratch and getting it FDA cleared. It covers the four phases of the process (Explore, Develop, Validate, and Document) and discusses the costs, time, and data requirements involved. It also provides advice on regulatory strategy, data annotation, and algorithm prototyping. If you're interested in developing a medical device involving AI/ML.

This article provides a questionnaire for those looking to get their AI algorithm FDA cleared. It covers topics such as validation studies, non-ML and ML software items, and acceptance criteria for the device. Reading this article can help uncover gaps that may need to be addressed before FDA submission.

This article outlines the challenges of bringing radiology AI/ML to market, from clinical workflow to IT installation. It provides insights and lessons learned for those interested in developing and deploying AI models in radiology.

FDA's Draft PCCP (Predetermined Change Control Plan) guidance should streamline model changes AI/ML-enabled devices, allowing for pre-approved modifications without the need for additional marketing submissions.

This webinar covers the four phases of the regulatory process: Explore, Develop, Validate, and Document. It also discusses the costs, time, and data requirements involved in the process. Additionally, it provides advice on regulatory strategy, data annotation, and algorithm prototyping.

This FDA guidance applies to AI/ML-enabled radiology software that assists in the detection of disease. Many of the principles and approaches, however, apply beyond this limited scope.

This guidance is describes FDA’s thoughts on evaluating the performance of CADe radiological devices. Designing standalone and clinical performance testing studies for AI/ML-enabled SaMD can be tricky. You want to do enough testing to ensure the device is safe and effective, yet you also need to avoid going bankrupt on the way!

This document outlines the technical performance assessment required for premarket submissions of quantitative imaging devices to the FDA. It provides guidance on the evaluation of the safety and effectiveness of these devices, including the methodology for quantitative imaging and the appropriate use of imaging biomarkers. Medical device manufacturers may be interested in reading this document to ensure their products meet the required standards and to understand the regulatory expectations for premarket submissions.

Introduces standardized terms and definitions to foster a consistent understanding across the industry for ML-enabled devices.

How can machine learning help us improve medical images for human viewers or image processors? This article describes one machine learning method (generative adversarial networks) that has been adopted by the medical imaging community to enhance medical images.

Want to know how we develop safe, effective, and FDA compliant machine learning algorithms? This article describes how we develop machine learning algorithms, points out common pitfalls, and makes documentation recommendations.

There are many approaches to choosing a medical imaging segmentation algorithm. In this article, we provide an overview of how to choose a neural network architecture for medical image segmentation.

Annotating medical images is time-consuming and expensive. In this article, we explain how self-supervised learning can stretch limited training data and compare it to transfer learning. We also explore three self-supervised learning medical imaging tasks.

Experienced radiologists can identify the anatomical location of an axial CT slice within a second. They may say the slice is “near the apex of the heart” or “at the C7 vertebrae.” These anatomical landmarks are difficult to describe or detect using manually created features, but neural networks excel at this sort of pattern recognition. Can we create a neural network capable of performing slice localization with similar speed and accuracy to a radiologist?

If you followed along with our last post, we developed a deep-learning model that achieves our goal of identifying Simpsons characters in an image. However, as with all software development tasks, getting a working program is only half the battle. In order to maintain a program and fix bugs, the developer must understand the system– in particular, they must understand how it fails as well as how it succeeds. This can be quite a difficult task for deep-learning models, as they are black-boxes by nature of their construction. However, there are some techniques we have at our disposal to open up the black box and get a view into what is happening in our trained model; these can help us to find “bugs” in the model’s learning and even indicate how to resolve them. Among the many techniques to visualize the internals of a deep learning model, we will be focusing on the use of class activation maps.