Medical image classification is a crucial component of AI-assisted diagnostics, and the accessibility of medical systems has garnered significant interest. Ensemble learning (EL) has shown considerable promise in the field of image analysis. However, to improve accuracy in medical image classification, most EL-based methods manually set the size of feature subsets and underutilize the characteristics of various medical images. Traditional EL-based methods also tend to repeatedly search a limited number of key features and select feature subsets separately for each classifier, resulting in low effective utilization of features. To address these issues, this paper proposes a self-adaptive collaborative EL (SCEL) model for medical image classification. Our approach introduces a self-adaptive feature selection task designed to identify optimal feature subsets of varying sizes for the corresponding classifiers. Additionally, these tasks collaboratively search for optimal feature subsets by sharing common knowledge, which enhances search efficiency and avoids redundant searches. To achieve self-adaptive feature selection and facilitate knowledge interaction, an evolutionary algorithm is introduced. Finally, by combining the size of feature subsets with classification accuracy, we propose a multi-criteria feature subset evaluation strategy to select feature subsets that demonstrate strong classification performance while remaining as compact as possible. Experimental results demonstrate that the proposed model can generate optimal feature subsets for each basic classifier and exhibits excellent classification performance across skin medical images.