ADOP-NAS: Evolutionary Neural Architecture Search for Digital Pathology Via Diverse Initialization and Adaptive Mutation

Abstract

Digital pathology requires accurate, efficient patch-level processing due to the gigapixel scale of Whole-Slide Images (WSIs). We propose an evolutionary Neural Architecture Search (NAS) framework tailored for WSI pipelines. We redesign the search space using 2D depthwise separable convolutions to minimize computational load and introduce a search strategy combining chaos-driven, opposition-based initialization with adaptive mutation to enhance exploration and prevent premature convergence. We evaluate the discovered models on four pathology benchmarks (EBHI, BCNB, TCGA, and SPIDER) under a unified training and WSI-level evaluation protocol, where slide-level predictions are obtained by majority voting over patch-level outputs. Our method discovers compact networks (as low as 0.01, MB in parameter storage size) achieving highly competitive or improved performance relative to CNN, transformer, and NAS baselines. For instance, our models reach ACC = 0.985 and F1 = 0.986 while being ∼ 44× smaller than ResNet-50. On challenging datasets, they improve F1 and recall while remaining lightweight, matching VGG-16 performance with ∼ 537× fewer parameters. Across all benchmarks, the proposed method achieves up to +20.2% improvement in F1 and ∼ 537× reduction in parameter count relative to strong baselines, demonstrating a favorable accuracyefficiency trade-off. Ablation studies indicate that the proposed initialization and mutation components significantly improve fitness (𝑝 < 0.05) without increasing FLOPs or parameters, while the selection mechanism shows a consistent but non-significant trend. The proposed framework provides a practical, resource-efficient approach for deploying accurate models in clinical digital pathology, addressing computational constraints and label scarcity, and lays the groundwork for future NAS research.