3144 - Preserving Boundary Accuracy in Compact 3D nnU-Net via Boundary-Privileged Knowledge Distillation

Purpose/Objective(s): Knowledge distillation enables fast 3D segmentation when full-scale models are impractical in resource and time-sensitive clinical workflows such as adaptive radiotherapy. However, conventional KD overemphasizes background and global similarity, degrading boundary and small-structure accuracy, that is critical in radiotherapy where millimeter-level contour boundary misses can materially change DVH-driven optimization under steep dose gradients and tight OAR constraints. To resolve this, we propose Boundary Saliency Injected Decoupled (BSID) Distillation, a distillation strategy that emphasizes boundary learning.

Materials/Methods: Our method first derives a teacher-based boundary importance map using a fixed 3D Prewitt edge operator, then enforces boundary-gradient alignment between teacher and student, and finally uses this map to weight only the target/ distillation term while keeping the background term global. We evaluated an nnU-Net teacher and compact student model under a fixed architecture. The student was trained with no distillation, logit-based KD, structured knowledge distillation (SKD), and intra-class feature variation distillation (IFVD), and our method, and tested on LiTS 2017 (liver tumor, n = 131) and KiTS 2023 (kidney tumor, n = 599).

Results: Table 1 summarizes the performance gains of student models with different distillation strategies. Compared with the supervised student baseline, our method improves LiTS 2017 DSC/NSD(mm)/HD95(mm) from 70.67/62.62/70.60 to 78.78/72.31/47.80, and improves KiTS 2023 from 68.65/57.72/168.44 to 84.61/74.52/35.07, outperforming logit-KD, SKD, and IFVD, which show limited or even negative gains on KiTS. Notably, these improvements are achieved without increasing model complexity: both proposed and baseline model uses ~75% fewer parameters than the teacher model.

Conclusion: BSID improves boundary fidelity and overall segmentation performance in compact 3D nnU-Net students without adding inference-time overhead, supporting efficient deployment in boundary-sensitive, time-critical clinical workflows such as online adaptive radiotherapy. Moreover, BSID operates entirely in output space as architecture-agnostic and can be applied to other 3D segmentation backbones.