TLDR: MAIA is an open-source, Kubernetes-based platform designed to streamline the development, deployment, and integration of AI in healthcare. It fosters collaboration between clinicians, researchers, and AI developers by providing a comprehensive suite of tools for data management, model training, annotation, and clinical feedback. The platform’s modular architecture supports various workflows, including AI development, clinical data integration, and active learning, and has been successfully implemented at KTH Royal Institute of Technology and Karolinska University Hospital for real-world medical imaging AI applications.
The integration of Artificial Intelligence (AI) into healthcare promises to transform patient care and optimize clinical decision-making. However, bridging the gap between cutting-edge technical innovation and practical healthcare applications requires robust collaborative platforms. This is where MAIA, the Medical Artificial Intelligence Assistant, steps in.
MAIA is an open-source platform specifically designed to foster interdisciplinary collaboration among clinicians, researchers, and AI developers. Built on the powerful Kubernetes infrastructure, MAIA offers a modular and scalable environment. It comes equipped with integrated tools for every stage of the AI lifecycle, including data management, model development, annotation, deployment, and crucial clinical feedback. Key features that make MAIA stand out include project isolation, automated continuous integration and continuous deployment (CI/CD), and seamless integration with high-computing infrastructures and existing clinical workflows.
The platform supports real-world use cases in medical imaging AI, with successful deployments in both academic and clinical environments. By promoting collaboration and interoperability, MAIA aims to accelerate the translation of AI research into impactful clinical solutions, all while emphasizing reproducibility, transparency, and a user-centered design.
Core Principles and Architecture
MAIA’s core principles are centered around enhancing AI education for healthcare professionals, integrating AI advancements with medical research, and deploying AI solutions directly into clinical workflows to improve patient care and operational efficiency.
Underpinning MAIA’s capabilities is its Kubernetes-based architecture, which ensures scalability, security, and efficient resource management. It supports various Kubernetes distributions and employs a “Federation of Clusters” architecture, allowing deployment across multiple network-independent infrastructures. This design facilitates inter-connectivity between isolated clusters and flexible resource allocation. The platform also integrates ArgoCD for CI/CD practices, ensuring consistent management and automated updates of deployed modules.
To manage multiple projects, MAIA utilizes “MAIA Namespaces,” which are isolated virtual environments built on Kubernetes namespaces. Each namespace hosts individual projects, ensuring that all necessary tools and resources are available within its scope while remaining independent from other projects.
Integrated Tools and Modules
Each MAIA namespace is equipped with a comprehensive suite of integrated tools:
-
MAIA Workspace: The central hub, built on JupyterHub, offering isolated workspaces with allocated resources (including GPUs) for AI model development. It provides access points like Jupyter, remote desktop, and SSH, along with scientific computing environments such as Visual Studio Code, MATLAB, and RStudio.
-
MLFlow: For logging and monitoring machine learning experiments and serving as a registry for trained models.
-
MinIO: A cloud storage service for file management and sharing.
-
Orthanc and OHIF Viewer: An open-source DICOM PACS for storing and sharing medical images, integrated with a web-based DICOM viewer.
-
KubeFlow: A platform for building and deploying portable machine learning workflows.
-
XNAT and MONAI Deploy: Tools for integrating AI-based applications and automated AI-powered pipelines into clinical deployment scenarios.
-
Label Studio: An application for annotating various types of data, including medical images.
Beyond the user-facing namespaces, MAIA’s architecture includes internal layers: MAIA Core (for foundational deployment, networking, monitoring, and security) and MAIA Admin (for user and project management, authentication, and cross-cluster control).
High-Performance Computing and GPU Management
Recognizing the demanding computational needs of deep learning models and large 3D medical imaging datasets, MAIA includes a dedicated submodule called MAIA-HPC. This module simplifies data transfer, code management, and job submission to generic High-Performance Computing (HPC) systems, allowing users to leverage powerful external resources like those provided by the National Academic Infrastructure for Supercomputing in Sweden (NAISS).
To optimize resource utilization within MAIA clusters, a GPU booking system has been implemented. This allows users to reserve GPUs for specific periods, ensuring efficient and equitable distribution of these valuable computational resources.
Real-World Workflows and Implementations
MAIA’s modular design enables researchers and clinicians to define and adapt workflows for various scenarios. The platform supports a full AI development workflow, from data transfer via MinIO and preprocessing with Kubeflow, to model training monitored by MLFlow and prediction analysis using 3D Slicer.
A significant aspect is its integration into clinical environments. At Karolinska University Hospital (KUH), a workflow called RADIANCE facilitates the secure and pseudonymized export of medical DICOM images from clinical servers into MAIA’s Orthanc instance. This enables clinicians to visualize and process images using MAIA’s tools.
The platform also supports an active learning workflow, where radiologists actively validate AI-generated predictions and refine annotations using MONAI Label. This iterative feedback process allows models to be continuously retrained and improved based on expert clinical input.
MAIA has been successfully implemented at KTH Royal Institute of Technology, supporting over 20 researchers and playing a crucial role in graduate education by providing students with dedicated GPU resources and workspaces. At Karolinska University Hospital, MAIA serves as a central catalyst for AI model development, evaluation, validation, and clinical deployment. Notable applications include the detection and segmentation of vertebra metastasis in CT images and brain metastasis segmentation in MRIs, showcasing MAIA’s ability to handle complex medical imaging tasks and integrate AI models into clinical practice.
The MAIA platform is open-source and available on GitHub, with comprehensive documentation and an online instance. For more details, you can refer to the original research paper.
Also Read:
- MedIQA: A New Foundation Model for Medical Image Quality Assessment
- Pancreatic Cyst Analysis Enhanced by Fine-Tuned AI Models
Impact and Future Outlook
MAIA’s core mission is to bridge the gap between successful AI research and its real-world clinical application by placing clinicians and radiologists at the center of the AI lifecycle. It aims to deliver all necessary tools in compliance with healthcare security standards, ensuring the safe handling of sensitive data. While the path to wider adoption presents challenges, MAIA invites the medical AI community, healthcare institutions, and open-source contributors to collaborate in advancing AI-driven clinical care, ultimately accelerating the integration of AI into everyday practice and translating collaborative research into meaningful patient outcomes.
