Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation
Researchers have introduced CardiacNAS, a novel evolutionary neural architecture search (NAS) framework designed to improve cardiac magnetic resonance (CMR) segmentation. Addressing challenges such as low tissue contrast and fuzzy boundaries in CMR images, this method couples a UNet-like supernet with a specialized search space that includes depth, width, attention mechanisms, and residual scaling. The framework is explicitly resource-aware, jointly optimizing accuracy metrics like the Dice similarity coefficient (DSC) and Hausdorff distance (HD95) against model size and computational costs (FLOPs). Evaluated on the ACDC dataset, CardiacNAS outperforms six state-of-the-art methods, achieving a 93.22% average DSC and 4.73 mm HD95 with only 3.58 million parameters and 14.56 GFLOPs. The study highlights that specific architectural choices, particularly in attention and fusion, significantly enhance boundary fidelity. This approach offers a principled, efficient solution for deployable medical image analysis, balancing high performance with transparent reporting of computational complexity.
Wire timeline
Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation
Researchers have introduced CardiacNAS, a novel evolutionary neural architecture search (NAS) framework designed to improve cardiac magnetic resonance (CMR) segmentation. Addressing challenges such as low tissue contrast and fuzzy boundaries in CMR images, this method couples a UNet-like supernet with a specialized search space that includes depth, width, attention mechanisms, and residual scaling. The framework is explicitly resource-aware, jointly optimizing accuracy metrics like the Dice similarity coefficient (DSC) and Hausdorff distance (HD95) against model size and computational costs (FLOPs). Evaluated on the ACDC dataset, CardiacNAS outperforms six state-of-the-art methods, achieving a 93.22% average DSC and 4.73 mm HD95 with only 3.58 million parameters and 14.56 GFLOPs. The study highlights that specific architectural choices, particularly in attention and fusion, significantly enhance boundary fidelity. This approach offers a principled, efficient solution for deployable medical image analysis, balancing high performance with transparent reporting of computational complexity.
cs.AI updates on arXiv.org