Source-linked AI summary
LightMedSeg-ISLES: Stroke Lesion Segmentation with 81x Fewer Parameters than nnU-Net
Giorgi Nikvashvili, Hanxue Gu, Jie Bao, Kang Wang, Yang Yang
TL;DR
Large networks and ensembles perform strongly in stroke-lesion segmentation but impose storage and inference burdens. LightMedSeg-ISLES uses a 1.26-million-parameter compact pipeline with a tailored training and inference recipe, retaining most of filtered nnU-Net’s Dice while improving lesion-wise F1 and reducing model and patch-level compute demands.
Problem
Large segmentation networks and ensembles create storage and repeated-inference demands, motivating compact alternatives for T1-weighted stroke-lesion segmentation.
Method
LightMedSeg-ISLES combines a compact backbone with lesion-aware sampling, deep supervision, strong augmentation, and native-space restoration.
Results
LightMedSeg retains 97.5% of filtered nnU-Net’s Dice, improves lesion-wise F1 by 0.055, and slightly exceeds filtered UNETR++ and nnFormer.
Takeaways & Limitations
A 1.26-million-parameter single checkpoint provides a practical alternative to much larger segmentation systems, using 81.4× fewer parameters and 4.7× fewer operating-point FLOPs per standardized patch.
Takeaways & Limitations
Reported operating points use different selection protocols, and whole-volume latency depends on image size and hardware despite standardized patch-level efficiency reporting.
Abstract
from arXiv · showhide
Large networks and ensembles often lead medical image segmentation challenges, but their storage and inference demands complicate deployment. We present LightMedSeg-ISLES, a 1.26-million-parameter pipeline for T1-weighted stroke lesion segmentation in ISLES'26. On a 146-case held-out cohort, flip test-time augmentation produces 0.618 mean Dice and 0.599 lesion-wise F1. A 102.35-million-parameter nnU-Net ResEnc-L produces 0.634 Dice and 0.544 lesion-wise F1 after size filtering. LightMedSeg therefore retains 97.5\% of nnU-Net's Dice with 81.4$\times$ fewer parameters while improving lesion-wise F1 by 0.055. Its four-pass TTA operating point requires 4.7$\times$ fewer FLOPs per standardized patch than nnU-Net. It also slightly exceeds filtered UNETR++ and nnFormer. Longer training and stronger augmentation add 0.0358 Dice without increasing capacity, establishing a strong single-checkpoint alternative to much larger models.
1 Introduction
LightMedSeg-ISLES targets stroke-lesion segmentation with a compact single model designed to preserve the accuracy of much larger systems while reducing deployment demands.
- 1 Introduction: Large networks and ensembles create storage and repeated-inference demands in medical image segmentation challenges.The cited examples include a 25-model nnU-Net ensemble and a large residual nnU-Net.
- 1 Introduction: The method combines a few-parameter backbone with local structural priors, spatial anchors, lesion-aware sampling, deep supervision, and augmentation.These components target local detail, global context, and heterogeneous T1-weighted MRI.
- 1 Introduction: 1.26 million parameters let LightMedSeg retain 97.5% of filtered nnU-Net’s Dice and improve lesion-wise F1 by 0.055.The pipeline compares against nnU-Net and additional large baselines.
2 Methods
The method uses a compact 3D encoder–decoder with spatial anchors, local structural priors, multi-scale fusion, and training and inference procedures tailored to heterogeneous stroke MRI.
- 2.1 Architecture: LightMedSeg is a compact 3D U-Net-like encoder–decoder designed for a high performance-to-parameter ratio.Its encoder uses staged channels, multi-scale skip fusion, and deep supervision with only the main head used for inference.
- 2.1 Architecture: Spatial anchors, adaptive feature mixing, and multi-scale skip fusion combine global context with local detail before decoding.The LSPM also generates an updated embedding through structurally weighted expert projections.
- 2.1 Architecture: The local structural prior module produces a texture map from smoothed-versus-input features, enabling the encoder to preferentially process high-detail regions.The map uses a pointwise convolution and sigmoid; values near 1 indicate high-detail regions and values near 0 indicate low-detail regions.
- 2.3 Variants: A boundary-refined variant adds a full-resolution convolution after concatenating the upsampled base output with the original input.The study also evaluates a padded 256^3 input for this variant.
- 2.2 Preprocessing and Training: The pipeline combines N4 correction, percentile clipping, z-score normalization, RAS reorientation, isotropic resampling, cropping or padding, and deep supervision.Inference uses overlapping windows with Gaussian blending, four-pass flip TTA, thresholding, component filtering, and native-grid restoration.
- 2.2 Preprocessing and Training: Training uses lesion-centered and background-centered 128^3 patches, with augmentation including flips, deformation, resolution changes, smoothing, noise, and intensity shifts.Half of the patches are centered on lesions to increase exposure to small lesions.
3 Experiments and Results
On the 146-case held-out cohort, LightMedSeg’s final TTA operating point approaches filtered nnU-Net overlap while using far fewer parameters and FLOPs, and exceeds the evaluated transformer baselines. Training and lesion-level analyses show gains from stronger optimization and fewer false positives, with remaining differences concentrated by lesion size.
- 3.3 Training Recipe: 0.0358 Dice is added at fixed 1.26-million-parameter capacity through longer training and stronger augmentation, rising from 0.5705 at 100 epochs to 0.6063 after convergence with aug-v2.The compared variants use nine LightMedSeg designs trained for 100 epochs with deep supervision and AdamW.
- 3.4 Comparison with Larger Models: 0.6178 Dice and 0.5992 lesion-wise F1 are achieved by LightMedSeg with 1.26 million parameters, versus 0.6339 Dice and 0.5441 F1 for filtered nnU-Net with 102.35 million parameters.This corresponds to 97.5% of nnU-Net’s Dice, 81.4× fewer parameters, and a 0.0551 F1 improvement.
- 3.4 Comparison with Larger Models: 0.6464 is the raw nnU-Net Dice, but size filtering reduces Dice to 0.6339 while increasing lesion-wise F1 from 0.5317 to 0.5441.This contrast motivates reporting both voxel overlap and lesion detection.
- 3.4 Comparison with Larger Models: 0.6166/0.5782 and 0.6142/0.5742 Dice/F1 are obtained by filtered UNETR++ and nnFormer, respectively, while LightMedSeg exceeds both with substantially fewer parameters.LightMedSeg uses 33.9× fewer parameters than UNETR++ and 119.2× fewer than nnFormer.
- 3.4 Comparison with Larger Models: 305.9 GFLOPs for four-pass TTA are required by LightMedSeg, 4.7× below nnU-Net’s 1434.8 GFLOPs and 1.6× below nnFormer’s 486.1 GFLOPs.A single LightMedSeg pass requires 76.5 GFLOPs on the common 128^3 patch.
- 3.5 Lesion-Level Analysis: 59 of 146 case comparisons favor LightMedSeg over nnU-Net, with paired Wilcoxon p = 0.095; Q2 contains the largest Dice gap, while Q4 models are close.Q2 values are 0.588 versus 0.645, and Q4 values range from 0.797 to 0.814 Dice.
- 3.5 Lesion-Level Analysis: 234/84/229 TP/FP/FN counts after TTA replace 235/107/228, so the F1 gain comes mainly from fewer false positives.The comparison concerns lesion counts under TTA.
4 Discussion
The discussion argues that LightMedSeg’s results come from training strategy rather than added capacity, while identifying small-to-medium lesions as the main remaining gap. It also notes that operating points use different selection protocols across model families.
- 4 Discussion: 0.0358 Dice is added by longer training and aug-v2 without changing network capacity, while boundary refinement does not improve the base design.The discussion attributes final performance to training strategy rather than added capacity.
- 4 Discussion: Small-to-medium lesions contain the remaining performance gap between LightMedSeg and the larger comparison model.This is the lesion-size localization reported for the final model.
- 4 Discussion: LightMedSeg and nnU-Net operating points use the internal cohort, whereas transformer thresholds come from the validation split.Results are computed in 1-mm analysis space, and efficiency uses standardized patch-level FLOPs rather than whole-volume latency.
5 Conclusion
LightMedSeg-ISLES presents a compact single-checkpoint alternative to large segmentation systems, retaining near-baseline Dice while reducing model size and operating-point computation.
- 5 Conclusion: 97.5% of filtered nnU-Net’s Dice is retained by LightMedSeg-ISLES with 81.4× fewer parameters and 4.7× fewer operating-point FLOPs per standardized patch.The conclusion frames the compact backbone and training recipe as a practical alternative to larger systems.