Digital Breast Tomosynthesis (DBT) improves breast cancer diagnosis by providing 3D images of the breast. Many studies develop AI models to improve DBT interpretation, but current methods often rely on traditional bounding boxes fusion techniques, which are not well-suited for DBT images. To address these challenges, this study presents DBT-WBF, a novel approach for fusing boxes across slices in DBT images. DBT-WBF builds upon the Weighted Boxes Fusion with several improvements. Specifically, it limits the process of boxes fusion to a reasonable slice range and incorporates a slice discrepancy weight for more reliable fusion. Additionally, DBT-WBF employs a neighborhood consistency filter to eliminate noisy boxes and improve lesion slice selection accuracy. DBT-WBF was evaluated on the BCS-DBT dataset and outperformed existing fusion models with an average sensitivity of 94.21% (95% CI: 90.13–97.69). DBT-WBF can be easily integrated into other frameworks without retraining. The code is available at: https://github.com/jjjjjjjjj58/DBT-WBF .