MLXOps4Medic: A Service Framework for Machine Learning and Explainability Operations in Medical Imaging AI Development

Integrating AI models with the medical domain is challenging as it involves more complex workflows compared to traditional machine learning operations (MLOps). Requirements such as model performance monitoring, calibration, and retraining on live medical data, as well as underdeveloped XAI and MLOps integration, have not yet been fully covered by existing works. This paper proposes the MLXOps4Medic, an initial machine-learning-and-explaining operations (MLXOps) framework based on the microservices architecture for AI model development in the medical domain. The framework focuses on four types of tasks in the MLXOps: training, model evaluation, XAI explanation generation, and XAI evaluation. It can orchestrate tasks between different types of microservices, automatically transferring the necessary data and configurations for task execution to reduce operational overhead. Additionally, the framework collects the provenance of the workflow metadata and builds a provenance network, which facilitates workflow tracing and reproducibility throughout the complex MLXOps lifecycle. Three case studies illustrate the key findings concerning the framework: (1) a significant reduction in operational overhead by approximately 72.55% resulting from complex AI experiments; (2) facilitation of MLXOps customization for complex workflows such as data drift monitoring, integration of multimodal medical AI models, and reproduction of historical training data; and (3) provision for prediction and explanation generation for medical AI models through a web portal. Furthermore, the framework is compared with the well-established MLOps framework called MLflow on a general domain MLX pipeline execution, demonstrating observable efficiency in terms of zero-code configuration for the training task and utilizing 36.14% less time for training task execution. These key findings demonstrate that the MLXOps4Medic framework delivers an efficient and highly adaptable MLXOps solution for AI and medical integration. By addressing identified challenges in workflow orchestration, it serves as a foundational framework with far-reaching implications for future research and applications across diverse domains.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.