Multiple teachers-meticulous student: A domain adaptive meta-knowledge distillation model for medical image...

Deep learning (DL) has shown to be a powerful tool in medical image classification, yet it still faces major practical challenges. These include performance degradation due to domain shift, dependency on large-scale annotated datasets, high model complexity that limits deployment, and concerns regarding patient data privacy.

To address these challenges, we propose a framework called Multiple Teachers-Meticulous Student (MT-MS) for source-free domain adaptive medical image classification using meta-knowledge distillation.

In the proposed method, multiple teacher models trained on heterogeneous source domains are used to guide a compact student model without needing access to the source data. This source-free setup preserves privacy only by learning from parameters of teacher models. To enable robust adaptation to new target domains, the student learns to fuse and refine the knowledge from multiple teachers using a meta-learning strategy. The architecture integrates convolutional and attentive components to maximize knowledge transfer, and an adaptive training schedule balances teacher guidance during the learning process.

We evaluate the proposed method on six public medical imaging datasets spanning various imaging modalities and clinical applications. The experiments focus on a binary classification task of respiratory motion artifact detection as a representative challenge. The results demonstrate that the MT-MS framework achieves strong performance in accuracy and F1-score while maintaining a lightweight structure suitable for deployment in real-world clinical environments.

Overall, MT-MS provides a unified solution for source-free domain adaptation, data-efficient training, knowledge consolidation, and privacy preservation, making it a promising approach for practical medical image classification across domains.