Individually Harmful, Jointly Beneficial: Compression Strategy Interactions in Brain Tumor Segmentation

Jiawei Chen, Qizhe Zhang, Hongsheng Chang, Yuzhou Chen, Guoxi Zhang, Naohao Huang, Shanghang Zhang, Xueqing Yan

Keywords: model compression, brain tumor segmentation, interaction effects

Registration Requirement: Yes

Abstract:

A counter-intuitive finding is reported in model compression for 3D brain tumor segmentation: two strategies that individually degrade or barely affect performance, Mamba block relocation and hybrid-score token pruning, recover near-baseline accuracy when combined under $2.2{\times}$ parameter reduction.
In experiments on SegMamba-V2 (BraTS 2023), Mamba relocation alone decreases Mean Dice by 2.45 percentage points at channel width $C{=}32$, yet the addition of token pruning reverses this loss entirely, achieving 91.23\% Mean Dice (vs. 91.31\% uncompressed baseline with 60.69M vs. 138.78M parameters).
This positive interaction ($+$2.92 Dice points beyond additive expectation) is observed only at reduced capacity ($C{=}32$) and is absent at $C{=}48$ ($-$0.32), suggesting a state capacity bottleneck in which token pruning compensates for Mamba's difficulty processing high-resolution background tokens under limited channel width.
The interaction is strongest for the Enhancing Tumor class ($+$3.91 points), the smallest and most boundary-sensitive target.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 133