Detail Consistent Stage-Wise Distillation for Efficient 3D MRI Segmentation

2026-05-25

AI Summary

This work addresses the performance degradation of compressed 3D medical image segmentation models, which often arises from the loss of fine-grained structures—such as small lesions and sharp boundaries—due to multi-stage downsampling. To mitigate this issue, the authors propose a Detail-Consistent Distillation (DCD) framework that performs hierarchical wavelet-domain distillation between teacher and student encoder features at each level during training. The method emphasizes preserving directional high-frequency detail components while relaxing constraints on low-frequency approximation components to avoid over-regularization of global semantics. Notably, DCD introduces no additional inference overhead and is compatible with mainstream architectures such as nnU-Net. Experimental results on the BraTS 2024 and ISLES 2022 benchmarks demonstrate significant improvements in segmentation accuracy for compressed models, and the implementation has been publicly released.

Abstract

Deploying high-performing 3D medical image segmenters (e.g., nnU-Net) is often limited by memory footprint and inference latency. Compression is therefore necessary, but compact 3D encoders tend to lose fine structural cues (small lesions and sharp boundaries) as downsampling repeats across multi-resolution stages. We propose Detail Consistent Distillation (DCD), a stage-wise distillation framework that preserves structural detail across scales by aligning teacher-student features in a wavelet-decomposed representation. At each encoder stage, DCD distills directional detail components in the wavelet domain while leaving the coarse approximation comparatively unconstrained, avoiding over-regularization of global semantics. DCD is used only during training and introduces no inference-time overhead. Experiments on the BraTS 2024 and ISLES 2022 benchmarks demonstrate that our approach achieves superior performance in MRI segmentation using 3D multi-modal data. Code and implementation details for DCD are publicly available at https://github.com/ClinicaAlpha/DCD-3D-MedSeg.

Problem

• 3D MRI segmentation
• model compression
• structural detail loss
• memory footprint
• inference latency

Innovation

• stage-wise distillation
• wavelet decomposition
• detail preservation
• 3D medical image segmentation
• model compression