R-PADE: Ranking-Guided Evolutionary Neural Architecture Search forMedical Image Segmentation

Neural architecture search (NAS) can automate medical image segmentation network design, but costly candidate evaluation limits the supervision available for evolutionary selection. Since replacement decisions depend on relative candidate quality, learning architecture preferences directly can provide guidance under constrained training budgets. In this paper, we propose Ranking-Aware Predictor-Assisted Differential Evolution (R-PADE) for U-Net-style architecture search. Specifically, a predictor trained with pairwise ranking supervision and contrastive regularization directly determines trial–target replacement in discrete differential evolution. Aggregated pairwise preferences prioritize candidates for proxy training, whose observations subsequently update the predictor. This connects evolutionary selection with selective candidate evaluation, reducing reliance on direct training for replacement decisions. Structured encoding and feasibility repair maintain valid architectures during the search. Experiments on CHAOS, BUSI, and ISIC demonstrate competitive segmentation performance with compact architectures. Compared with EU-Net, R-PADE reduces search GPU time by 35.6% while achieving higher Dice scores on CHAOS and BUSI. The ablation experiments compare the implemented predictor formulations and evaluate the contribution of contrastive regularization using matched search settings and proxy-evaluation budgets.