Enhancing 3D multi-organ segmentation via uncertainty guidance and boundary knowledge distillation

We propose the Uncertainty Guidance and Boundary Knowledge Distillation (UGBKD) framework for enhancing 3D multi-organ segmentation performance of student networks. UGBKD integrates three strategies: uncertainty-guided knowledge distillation, learning difficulty mining mechanism, and boundary knowledge distillation. The teacher-student distillation is adeptly guided by leveraging estimated uncertainty and the learning difficulty mining mechanism. Boundary knowledge distillation further alleviates blurred boundary challenges. Initially, a pre-trained denoising autoencoder DAE with anatomical perception priors is employed to estimate prediction uncertainty, and the uncertainty guided strategy promotes consistent knowledge transfer from the teacher. Subsequently, the learning difficulty mining mechanism focuses on difficult areas for the student. Lastly, boundary knowledge distillation extracts and transfers crucial boundary information to enhance the student's boundary perception. Extensive experiments on WORD and BTCV datasets validate our proposed method's effectiveness in improving segmentation accuracy and robustness. Code is available at https://github.com/wutianqi-Learning/UGBKD.