CardiacNAS: Resource-Aware Neural Architecture Search for Heart MRI Segmentation
A recent study introduces CardiacNAS, a new evolutionary framework designed for neural architecture search specifically for segmenting cardiac magnetic resonance (CMR) images. This framework utilizes a supernet that resembles UNet and focuses on a specialized search space for cardiac applications, which includes various parameters like depth, width, kernel size, and more. The goal is to improve segmentation accuracy, assessed through the Dice similarity coefficient and the 95th percentile Hausdorff distance, while also optimizing computational efficiency based on model size and FLOPs within set compute limits. Candidate architectures evolve through methods like crossover and mutation, and the findings were benchmarked against six top competitors on the ACDC dataset. The paper is available on arXiv under the identifier 2605.08238.
Key facts
- CardiacNAS is an evolutionary neural architecture search framework for CMR segmentation.
- It uses a UNet-like supernet with a cardiac-aware search space.
- Search space includes depth, width, kernel size, filter size, attention, fusion, activation, dropout, and residual scaling.
- Optimizes Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95) versus model size and FLOPs.
- Candidate architectures are trained with proxy budgets and evolved via crossover, mutation, and elitist selection.
- Evaluated on the ACDC dataset.
- Compared against six state-of-the-art methods.
- Paper available on arXiv:2605.08238.
Entities
Institutions
- arXiv