Lightweight Ensemble Learning based on Post-Hoc Probability Fusion with Its Application to Lung CT Image Classification.

INTRODUCTION

Deep learning model fusion has become one of the most effective methods to improve the performance of ensemble learning frameworks. However, the optimization of critical parameters within ensemble learning models to improve prediction accuracy remains a challenging issue. This study proposes a lightweight ensemble learning model fusion method based on post-hoc weighted averaging of predicted probabilities.

METHODS

By appropriately adjusting the fusion weight coefficient and the decision threshold, the optimal fusion strategy is investigated to improve performance metrics including accuracy, precision, recall, and F1 score. To verify the effectiveness of the proposed lightweight ensemble learning fusion method, it is applied to a medical classification task using lung computed tomography (CT) images, which aims to distinguish between pneumonia and lung cancer, two diseases with inter-class visual similarity. We construct a dataset containing 859 lung CT images representing pneumonia and lung cancer. First, we evaluated nine deep learning architectures from the VGG family (VGG11, VGG13, VGG16, VGG19) and the ResNet family (ResNet34, ResNet50, ResNeXt50, ResNet101, ResNeXt101). The results show that ResNet-based models consistently outperform VGG-based models. Therefore, we adopt the proposed lightweight weighted probability fusion strategy to integrate two complementary ResNet models, namely ResNet34 and ResNeXt50.

RESULTS

Experimental results show that the optimized fusion model achieves an accuracy of 96.67%, precision of 94.68%, recall of 98.88%, and F1 score of 96.74%, outperforming the individual models across all evaluation metrics.

DISCUSSION: These results show that the proposed lightweight ensemble learning method can enhance classification accuracy and decision robustness. This suggests that weighted probability fusion is effective for tasks with inter-class visual similarity.

CONCLUSION

The proposed method provides a feasible strategy for a model ensemble framework in medical image analysis, particularly when balancing performance and computational efficiency.