Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation

Technical Summary: Resource-Aware Evolutionary Neural Architecture Search for Cardiac MRI Segmentation

Problem Statement
Cardiac magnetic resonance (CMR) segmentation is critical for the quantitative assessment of ventricular structure and function, specifically delineating the left ventricle (LV), right ventricle (RV), and myocardium (MYO). However, achieving reliable delineation is hindered by low tissue contrast, fuzzy boundaries (particularly around papillary muscles and insertion points), inter-patient variability, and domain shifts across different scanners and protocols. While deep learning, particularly encoder–decoder architectures like U-Net, has become the dominant paradigm, practitioners face significant trade-offs between boundary quality (measured by Dice Similarity Coefficient [DSC] and 95th-percentile Hausdorff distance [HD95]) and computational efficiency (parameters and FLOPs). Existing Neural Architecture Search (NAS) methods often require substantial engineering effort to balance these competing objectives or fail to explicitly optimize for deployable resource constraints in cardiac-specific contexts.

Methodology: CardiacNAS
The authors propose CardiacNAS, an evolutionary Neural Architecture Search framework designed to automate the design of efficient CMR segmentation models under explicit resource constraints. The framework operates through the following components:

1. Cardiac-Aware Supernet and Search Space:
   The backbone is a U-Net-like supernet structured into an encoder, bottleneck, and decoder with skip connections. Unlike standard NAS, the search space is explicitly tailored for cardiac imaging, parameterized by a genotype $\gamma$ that controls:

   • Structural Hyperparameters: Filter sizes ($fl$), kernel sizes ($kn$), and number of layers ($ns$).
   • Advanced Modules: Attention mechanisms (squeeze-excitation, self-attention), fusion/aggregation strategies (add, concat, weighted sum), and activation functions (ReLU, ELU, Tanh, Sigmoid).
   • Regularization and Scaling: Dropout rates ($dp$) and a continuous residual scaling factor ($\alpha$) applied to skip connections.
   • Input: Fixed-size normalized cardiac MR slices ($128 \times 128$).

2. Evolutionary Search Process:
   The framework employs an evolutionary algorithm (Algorithm 1) to optimize architectures:

   • Initialization: A population of candidate architectures is sampled from the search space.
   • Variation: Candidates undergo crossover (uniform exchange of genes, blending for continuous $\alpha$) and mutation (resampling discrete variables or jittering continuous ones).
   • Proxy Training & Evaluation: Each candidate is instantiated and trained with a limited budget (proxy training) on the training set. Fitness is evaluated on a validation set based on a multi-objective function balancing DSC and HD95 against model complexity (Parameters and FLOPs).
   • Selection: An elitist replacement strategy selects the best performers to form the next generation, maintaining an archive of non-dominated solutions.

3. Resource-Awareness:
   The search explicitly optimizes the trade-off between segmentation fidelity and computational cost, ensuring the final architecture is deployable within fixed compute budgets. The total search cost (including all candidate evaluations) is factored into the efficiency analysis.

Key Contributions
The paper outlines three primary contributions:

• Cardiac-Aware Supernet: A specialized search space integrating multi-scale context, attention, and flexible skip-fusion strategies specifically designed to target crisp LV/MYO/RV boundaries under tight complexity budgets.
• Resource-Aware Evolutionary NAS: A procedure that explicitly trades off segmentation quality (DSC/HD95) against model complexity (Parameters/FLOPs) using proxy training and evolutionary selection, eliminating the need for manual design balancing.
• Comprehensive Evaluation: A detailed analysis including qualitative comparisons, learning-curve analyses, and design-factor correlation studies. These analyses clarify how specific architectural choices (e.g., self-attention, weighted-sum fusion, residual scaling) influence boundary fidelity and stability.

Experimental Results
The framework was evaluated on the ACDC dataset (150 patient cases) and compared against six state-of-the-art methods (U-Net, nnU-Net, TransUNet, BiX-NAS, HCT-Net, M-ENAS).

• Performance: CardiacNAS achieved an average DSC of 93.22% and an HD95 of 4.73 mm. This represents a 2.63% improvement in DSC over the next best evolutionary method (M-ENAS) and a 19.2% reduction in HD95.
• Efficiency: The resulting model is highly compact, requiring only 3.58M parameters and 14.56 GFLOPs.
• Search Cost: The entire NAS process (20 generations, 200 candidates) required only 0.18 GPU days, significantly lower than manual baselines and other evolutionary approaches.
• Statistical Significance: T-tests confirmed CardiacNAS achieved the highest statistical score ($p=0.029$) compared to baselines.
• Design Insights: Correlation analysis revealed that self-attention modules, weighted-sum fusion, and sigmoid activation functions positively correlated with higher DSC and lower HD95. The optimal configuration included a filter size of 96, kernel size of 3, 3 layers, 0.3 dropout, and a residual scaling factor of 0.4.
• Qualitative Analysis: Visual comparisons showed CardiacNAS produced crisper endocardial and epicardial boundaries and reduced RV boundary leakage compared to competitors, which often exhibited under/over-segmentation near the septum or cavity erosion.

Significance and Claims
The paper claims that CardiacNAS offers a principled, resource-aware approach to deployable CMR segmentation. By coupling a U-Net-like supernet with a cardiac-specific search space and evolutionary optimization, the framework successfully automates the design of models that achieve high accuracy without sacrificing computational efficiency. The authors emphasize the transparency of their approach, providing clear reporting of architectural complexity and compute budgets. The work demonstrates that searched attention mechanisms, fusion strategies, and residual scaling are critical factors in improving boundary fidelity and stability within compact compute budgets.

The paper concludes modestly, noting that future work will focus on extending the framework to 2.5D/3D representations to incorporate through-plane spatial context, evaluating on larger multi-center cohorts, refining proxy-training strategies, and integrating hardware-aware objectives such as latency and energy constraints.