Lightweight Container Orchestration for Reproducible AI and Deep Learning Segmentation of Coronary Arteries and the...

Background: Deep learning segmentation of coronary CT angiography (CCTA) can support visualization, quantitative analysis, and patient-specific research workflows, but local deployment is often limited by the GPU configuration, dependency drift, heterogeneous operating systems, and limited DevOps resources. Objective: This study aimed to develop and evaluate a cross-platform orchestrator for the reproducible execution of containerized cardiovascular segmentation models without the need for Kubernetes. The segmentation models were used as representative demonstration workloads. Materials and Methods: A lightweight container orchestrator was developed using Podman and Podman Desktop Machine. The system provides a standardized REST API, asynchronous job execution, parallel and cascaded inference modes, persistent logging, and an external evaluation module. Two containerized nnU-Net v2 services were implemented for coronary artery and aortic segmentation from CCTA. Models were trained on the ImageCAS data and evaluated on an independent external cohort of 200 CCTA examinations. The performance was assessed using the Dice similarity coefficient (DSC) and intersection-over-union (IoU). System latency, throughput, and resource utilization were also measured. Results: The mean end-to-end latency on Linux was 52.3 ± 7.1 s with GPU acceleration and 74.1 ± 8.3 s in CPU fallback mode. Parallel execution increased throughput from 44.4 to 62.1 cases/hour, with the expected latency increase due to resource contention. As workload validation, the coronary and aortic nnU-Net services achieved DSC values of 0.93 and 0.95, respectively. Conclusions: The proposed orchestrator enables reproducible, standardized, portable deployment of containerized CCTA segmentation models in on-premises research environments, reducing operational barriers while supporting visualization, validation, and downstream applications.