State-of-the-Art Stroke Lesion Segmentation at 1/1000th of Parameters

TL;DR

The paper tackles the challenge of efficient whole-brain stroke lesion segmentation in MRI by reviving MeshNet and introducing a multi-scale dilation pattern that captures broad context without downsampling or skip connections. It demonstrates that a 10-layer, parameter-efficient MeshNet with adaptive dilation can segment $256^3$ volumes directly, achieving competitive performance with only a fraction of the parameters of state-of-the-art models on the ARC dataset. MeshNet-26 reaches a DICE of $0.876$ with $147{,}474$ parameters and MeshNet-16 reaches $0.873$ with $56{,}194$ parameters, placing on the Pareto frontier against much larger models such as U-MAMBA-BOT and MedNeXt-M. The results indicate strong efficiency–accuracy trade-offs, enabling potential deployment in resource-limited settings and browser-based tools like brainchop, with future work extending validation to additional datasets.

Abstract

Efficient and accurate whole-brain lesion segmentation remains a challenge in medical image analysis. In this work, we revisit MeshNet, a parameter-efficient segmentation model, and introduce a novel multi-scale dilation pattern with an encoder-decoder structure. This innovation enables capturing broad contextual information and fine-grained details without traditional downsampling, upsampling, or skip-connections. Unlike previous approaches processing subvolumes or slices, we operate directly on whole-brain $256^3$ MRI volumes. Evaluations on the Aphasia Recovery Cohort (ARC) dataset demonstrate that MeshNet achieves superior or comparable DICE scores to state-of-the-art architectures such as MedNeXt and U-MAMBA at 1/1000th of parameters. Our results validate MeshNet's strong balance of efficiency and performance, making it particularly suitable for resource-limited environments such as web-based applications and opening new possibilities for the widespread deployment of advanced medical image analysis tools.

Figures (3)

• Figure 1: Comparison of DICE scores across models for whole-brain lesion segmentation. Models are ranked by median DICE score, with MeshNet-26 achieving the highest performance. Red asterisks (*) indicate models with statistically significant differences ($p < 0.05$, Holm-corrected Wilcoxon test) compared to MeshNet-26.
• Figure 2: Relationship between model complexity (represented as the the inverse of parameter count on a log scale) and median DICE score with interquartile range (IQR) error bars. Models with fewer parameters appear toward the right, while models with higher parameter counts are positioned on the left. MeshNet models are on the Pareto frontier indicating the balance between parameter efficiency and accuracy. The highest number of channels (MeshNet-26) produces the best DICE.
• Figure 3: Comparison of chronic lesion segmentation on T2-weighted MRI of the ARC dataset, with color-coded contours indicating each model's output. While most models display similar alignment with lesion boundaries, subtle differences are observed. MeshNet-16 and -26 generally show fewer issues with over- or under-segmentation. In contrast, other models, including MeshNet-5, MedNeXt-M, and U-MAMBA-BOT, exhibit more frequent deviations in specific regions and variability in capturing finer lesion details.