Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation

CardiacNAS presents an evolutionary neural architecture search framework that optimizes cardiac MRI segmentation models for both accuracy and computational efficiency. The approach achieves 93.22% dice similarity with only 3.58M parameters, demonstrating how resource-aware AI design can enable deployment of medical imaging models on resource-constrained environments.

CardiacNAS addresses a critical challenge in medical AI: deploying accurate segmentation models within strict computational budgets. Cardiac MRI analysis requires precise delineation of ventricular structures, but existing methods struggle with low tissue contrast and boundary ambiguity. The framework couples evolutionary algorithms with neural architecture search to automatically discover optimal model architectures rather than relying on manual design or resource-agnostic approaches.

The resource-aware optimization represents a meaningful shift in medical AI development. Rather than pursuing maximum accuracy without constraints, CardiacNAS jointly optimizes accuracy metrics (DSC and HD95) against model size and computational requirements. This pragmatic approach acknowledges that clinical deployment often requires edge-compatible or mobile-first solutions. The evolutionary search process—using crossover, mutation, and selection—proves effective at navigating the complex architectural space spanning depth, width, kernel configurations, attention mechanisms, and regularization strategies.

For the healthcare technology sector, this methodology has implications beyond cardiac imaging. The 3.58M parameter model with 14.56 GFLOPs computational demand makes high-quality segmentation accessible to hospitals with legacy infrastructure or resource-limited settings, potentially expanding diagnostic capability in underserved regions. The transparent reporting of architectural decisions and computational budgets also establishes reproducibility standards that regulatory bodies increasingly require for clinical AI systems.

The validation on the ACDC dataset and comparison against six state-of-the-art methods provides credible evidence of the approach's effectiveness. Future developments might extend this framework to other medical imaging modalities or explore federated learning scenarios where multiple institutions deploy similar architecture families with local data optimization.

• →Evolutionary NAS discovers cardiac segmentation models achieving 93.22% accuracy with only 3.58M parameters and 14.56 GFLOPs.
• →Resource-aware optimization jointly targets accuracy and computational efficiency rather than optimizing accuracy alone.
• →Attention mechanisms and residual scaling emerged as key architectural factors for improved boundary detection in cardiac imaging.
• →The framework enables deployment of clinical-grade segmentation on resource-constrained environments and edge devices.
• →Transparent reporting of model complexity and compute budgets establishes reproducibility standards for regulated medical AI systems.