Bronchiolitis obliterans (BO) is characterized by airflow obstruction resulting from inflammatory damage to the small airways in children. Current diagnostic methods heavily depend on clinical expertise, lacking efficient imaging-assisted tools. This paper introduces a novel deep learning framework based on ensemble learning to assist in the diagnosis of BO. Recognizing spatial feature similarities between BO and non-BO samples, an optimized network architecture has been devised to reveal high-order features. To addresss class imbalance, data redistribution strategies and grouped training methods have been proposed to ensure a balanced sample distribution. By integrating multiple independent models into an ensemble classifier, the classification accuracy has been significantly improved. Experimental results show that the single network model achieved an accuracy of 72.39% for BO and 82.81% for non-BO, which increase to 95.48% and 91.21% after model integration. These results underscore the efficacy of the proposed deep learning framework in pediatric BO classification, facilitating early diagnosis.