Structure-aware efficient compression for dental image segmentation using differentiable gates and masked knowledge distillation

초록

Recent advances in medical image segmentation using deep learning have developed dental technology. However, many dental clinics lack the minimum hardware infrastructure required to run high-performance deep learning algorithms in real-time, hindering their commercialization. Although filter pruning and knowledge distillation hold promise as resource-efficient solutions, they remain underexplored in medical/dental image segmentation. This paper presents a novel framework for compressing dental image segmentation networks. First, we efficiently explored sub-networks by removing unnecessary filters through a learnable differentiable gate. Second, during sub-network exploration, we further utilized the baseline network’s information through masked knowledge distillation. Through this approach, the proposed compression framework efficiently explores a more appropriate sub-network with minimal loss of the baseline network’s information for each structure. As a result, the proposed method was tested on dental anatomical structure segmentation on 3D CBCT and teeth segmentation on panoramic radiographs, achieving reductions in MACs by 90% and 95%, respectively, while showing only around a 1% decrease in performance based on the Dice score. The ability to achieve up to 95% MAC reduction with minimal Dice degradation highlights its potential for real-time deployment in resource-limited dental clinics, paving the way for practical clinical adoption. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2026.

키워드