DCD raises compressed 3D MRI segmentation mDice from 63.60% to 68.51% on BraTS 2024 and to 73.95% on ISLES 2022 while cutting parameters from 101.9M to 6.4M

A wavelet-domain subband selection strategy that partitions 3D DWT subbands into approximation A={LLL}, extreme high-frequency S={HHH}, and directional detail D={L,H}^3 \ {LLL,HHH}, distilling only D. Earlier frequency-aware distillation largely targeted 2D natural images with full-spectrum or whole-frequency matching; this work accounts for MRI spectral statistics by explicitly excluding the noise-dominated S band while leaving A unconstrained so global semantics are not over-regularized. Ablations show that on BraTS 2024 distilling S yields 61.98% mDice, below the 63.60% of no distillation, whereas distilling D yields 68.50%; on ISLES 2022 D reaches 73.95% versus 72.12% for S and 69.54% for A. Proposition 1 further argues, under white-noise and k-space energy-decay assumptions, that band S typically has lower subband SNR.

A stage-wise distillation framework that projects teacher and student features onto the detail subspace at each encoder stage and computes MSE after inverse-wavelet reconstruction back to the spatial domain, optimized jointly with the segmentation loss. Supervision is restricted to detail content while the alignment loss is computed in the same geometry as standard feature MSE, rather than directly on coefficient tensors whose layout depends on implementation choices; DWT/IDWT and the 1x1x1 projection are used only during training. Removing IDWT reconstruction degrades performance markedly: BraTS 2024 mDice drops from 68.51% to 62.87% and HD95 rises from 6.25 to 9.01; ISLES 2022 mDice falls from 73.95% to 70.06%.

Empirical validation of the accuracy-efficiency trade-off under compression on 3D brain tumor and stroke lesion segmentation benchmarks. With an nnU-Net topology and channel reduction factor r=4, DCD outperforms distillation baselines including CWD, IFVD, Logits, Feature, and FreeKD on both datasets. On BraTS 2024, DCD improves over the non-distilled student by +4.91% mDice (paired t-test p=8.64x10^-13, Wilcoxon p=4.40x10^-16) and over the previous best IFVD (66.54%) by +1.97% (p=0.0078 / 0.0080); on ISLES 2022 it reaches 73.95% mDice, +3.74% over the non-distilled student (p=0.0058 / 0.0008). Parameters drop from 101.945M (teacher) to 6.381M (student) and FLOPs from 19.505T to 1.277T on BraTS 2024.

Detail-sensitive regions benefit more: on BraTS 2024 the NETC subregion mDice rises from 41.26% to 54.36%. The gains are not uniform but concentrate in subregions tied to boundaries and thin structures, consistent with the authors' motivation that compression suppresses detail-sensitive representations. Table 2 reports a +13.10% NETC gain alongside SNFH 88.49%, ET 62.55%, and RC 68.62%; the qualitative comparison (Fig. 3) shows DCD predictions closer to ground truth.

This work targets researchers and engineering teams who need to deploy 3D MRI segmentation models on constrained hardware or in high-throughput settings: when the teacher follows nnU-Net topology, the student is built with channel reduction factor r=4, and teacher features are available during training, DCD serves as a plug-in training-time distillation term with no extra inference cost. The authors validate on BraTS 2024 (1,350 cases, four MRI modalities, four tumor subregions) and ISLES 2022 (250 cases, three modalities, a single lesion category), using the db4 wavelet, 3-level decomposition, and distillation weight 0.05. Code and implementation details are publicly available at https://github.com/ClinicaAlpha/DCD-3D-MedSeg.

A careful reader would still watch: band selection rests on the simplified white-noise and k-space energy-decay assumptions of Proposition 1, so whether S is always harmful under more complex real MRI noise and artifact structure remains open; whether the 3-level decomposition, db4 basis, and distillation weight 0.05 are robust across other datasets and resolutions; whether detail-band supervision still beats full-spectrum matching at compression strengths other than r=4; and whether the current validation, limited to two brain MRI benchmarks, transfers to other modalities, organs, and clinical workflows without independent evidence.