Skin pathological images contain essential diagnostic information across various scales. To effectively utilize multi-scale features, this study proposes a classification method based on multi-scale neural networks. The method involves a variable multi-scale neural network structure with a backbone network and multiple scale input branches inserted at different layers, facilitating feature extraction and fusion. Two search algorithms—a minimum cost-based search algorithm and a hill-climbing search algorithm—are introduced to identify the optimal network structure. Experimental results demonstrate that the proposed multi-scale network outperforms original networks in skin pathological image classification and that both search algorithms efficiently find near-optimal structures with reduced computational costs.