Responsible Deep Learning for Software as a Medical Device

TL;DR

This paper discusses the challenges, regulatory considerations, and evaluation strategies for deploying responsible deep learning models in medical imaging applications, emphasizing safety, fairness, and clinical integration.

Contribution

It provides a comprehensive overview of regulatory frameworks, evaluation methods, and opportunities for responsible AI deployment in medical device contexts, based on expert insights and case studies.

Findings

Performance of AI models varies across medical imaging modalities.

Regulatory evaluation is crucial for clinical deployment.

Strategies for addressing dataset biases are discussed.

Abstract

Tools, models and statistical methods for signal processing and medical image analysis and training deep learning models to create research prototypes for eventual clinical applications are of special interest to the biomedical imaging community. But material and optical properties of biological tissues are complex and not easily captured by imaging devices. Added complexity can be introduced by datasets with underrepresentation of medical images from races and ethnicities for deep learning, and limited knowledge about the regulatory framework needed for commercialization and safety of emerging Artificial Intelligence (AI) and Machine Learning (ML) technologies for medical image analysis. This extended version of the workshop paper presented at the special session of the 2022 IEEE 19th International Symposium on Biomedical Imaging, describes strategy and opportunities by University of California professors engaged in machine learning (section I) and clinical research (section II), the Office of Science and Engineering Laboratories (OSEL) section III, and officials at the US FDA in Center for Devices & Radiological Health (CDRH) section IV. Performance evaluations of AI/ML models of skin (RGB), tissue biopsy (digital pathology), and lungs and kidneys (Magnetic Resonance, X-ray, Computed Tomography) medical images for regulatory evaluations and real-world deployment are discussed.