Training-free multi-scale neural architecture search for high-incidence cancer prediction

Deep neural networks excel in high-incidence cancer prediction; however, designing networks that predict specific cancers is time-consuming and requires expert. The neural architecture search method offers a way to automate network design and has shown success in natural image. However, the small and varying lesion sizes in cancer image pose challenges, and most neural architecture search methods are computationally expensive and exhibit low agent correlation. Therefore, we propose a training-free multi-scale neural architecture search method for high-incidence cancer prediction. We introduce a multi-scale search space to address varying lesion sizes; and identify optimal scale combinations for feature extraction. To reduce computational costs and improve agent correlation, we design a training-free agent that evaluates network performance based on convergence, expressiveness, trainability, and complexity, enabling efficient neural architecture search implementation. Our extensive experiments on the NAS-Bench-201, MedmnistV2, LC25000, BreakHis, and CRC-5000 datasets show that our method outperforms both manually designed networks and state-of-the-art neural architecture search methods. The results demonstrate average improvements of 4.2%, 1.88%, 79.45%, 34.31%, and 31.71% in accuracy, area under the curve, search time, and Kendall and Spearman correlation coefficients, respectively.