Advancing Clinical AI Translation with High-quality Open Source Tools and Engineering Practice

This blog serves as a follow-up to our previous blog on confidently navigating Software as a Medical Device (SaMD) product development. In that blog, we explored the foundational principles and challenges of SaMD development. Here, we focus on the tools and processes we employ to ensure the development of high-quality AI components as part of SaMD products. Our approach leverages advanced open source tools and rigorous engineering practices, helping organizations build AI pipelines that meet FDA and clinical requirements.

High-Quality AI Software Development

Developing AI software for healthcare applications demands a heightened focus on quality. Unlike experimental AI models, clinical-grade AI systems must be:

• Robust and Reliable: Perform consistently under varied and unforeseen conditions.
• Traceable: Offer transparency into data handling, decision-making, and outcomes.
• Compliant: Meet regulatory requirements such as FDA
• Scalable: Adapt to different clinical settings and deployment environments.
• Maintainable: Support updates and continuous improvements without compromising safety or efficacy.

At Kitware, we understand these challenges and bring years of experience in developing open-source platforms and customized solutions for clinical-grade AI development. Our approach integrates the latest tools and practices to meet these stringent requirements.

Data Engineering

High-quality data pipelines are essential for building clinical-grade AI, encompassing robust data management, versioning, labeling, and exploration tools to ensure traceability and reproducibility. At Kitware, we use tools like DVC (Data Version Control) and Git-LFS to maintain a history of datasets, ensuring that every change is traceable and enabling the reproducibility of results. For data labeling and annotation, we leverage tools like MONAI Label, MIQA, and DIVE to enable domain experts to generate high-quality labeled datasets efficiently, particularly for medical imaging tasks. This ensures the creation of reliable datasets critical for AI model training. In the realm of data processing and exploration, we utilize tools such as DAGSTER to create robust data pipelines that manage dependencies and orchestrate complex workflows. Additionally, Python libraries like Pandas facilitate exploratory data analysis, providing actionable insights from healthcare datasets to inform AI development. These practices collectively ensure that data pipelines meet the stringent requirements of clinical-grade AI solutions.

AI Model Development

Our software engineering practices for experiments focus on robust workflows and tools. By integrating with Git, we ensure version control for code, providing full traceability of changes. As experts in MONAI, we develop and deploy flexible training frameworks, including distributed training setups, to accommodate the demands of advanced AI medical projects. Moreover, we utilize Weights and Biases to simplify hyperparameter tuning and optimize AI models to achieve clinical-grade performance, ensuring that the solutions we deliver meet the highest standards of quality and efficacy. In the realm of experiment management and training frameworks, we integrate MLflow to track experiments, manage models, and streamline transitions from development to deployment.

Deployment

Deployment is often the most critical phase for clinical-grade AI, and we ensure that AI models are scalable. For edge deployment, we optimize models using TensorRT, enabling real-time inference on edge devices in clinical settings. Additionally, we develop intuitive web interfaces using our advanced framework Resonant, allowing seamless interaction with AI models for clinicians and end-users. Our expertise in continuous integration and testing (CI/CD pipelines) automates testing, validation, and deployment processes, ensuring that updates maintain model integrity and performance without introducing risks.

Adopting Best Practices in Software Engineering

Traceability and conformance are essential in regulated environments, where every change in data, code, and models must be traceable to its origin. Kitware ensures comprehensive traceability, simplifying audits and regulatory submissions to meet stringent compliance requirements.

Our advanced data engineering practices provide robust foundations for building reliable AI models. Automated pipelines minimize errors, while versioning enables reproducibility, both of which are critical for research and clinical applications. These practices ensure that data integrity is maintained throughout the AI development lifecycle.

Rigorous testing and validation are integral to our processes. Kitware employs comprehensive testing protocols, including unit tests, integration tests, and system-level validations, to confirm that models meet predefined performance and safety thresholds. Tools such as CMake and CTest are used to streamline and automate these testing processes, ensuring high-quality outcomes that align with clinical and regulatory standards.

Why Choose Kitware for Clinical AI Development?

Kitware is uniquely positioned to help organizations navigate the complexities of AI-driven SaMD development:

• Open Source Leadership: As contributors to tools like MONAI, ITK, and VTK, Kitware is at the forefront of open source innovation.
• Custom Solutions: We tailor our services to meet the unique requirements of each organization and regulatory landscape.
• Collaborative Approach: Kitware fosters cross-disciplinary collaboration between engineers, clinicians, and regulators to define requirements and deliver solutions.
• Proven Expertise: With decades of experience in healthcare and software engineering, Kitware delivers results that are scalable, reliable, and compliant.

By combining these practices with powerful tools, Kitware empowers organizations to confidently engineer AI for SaMD, unlocking its potential to transform healthcare delivery and outcomes. Contact our team to learn more.