Research map / Detection and segmentation / CNN features
S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes S³-Mamba, a VMamba-based U-shape segmentation model with an enhanced VSS block, tensor-based cross-feature multi-scale attention, and regularized curriculum learning to improve small lesion segmentation.
Problem
Segmenting small lesions in medical images is difficult because they occupy a small image fraction and down-sampling erodes their local features, with training biased toward large objects.
Approach
Combines an EnVSS block (VSS block with DWConv removed, channel-wise SE-style reweighting, and two residual connections), a tensor-based attention over input, edge, and prediction features at multiple patch scales that modulates decoder features, and a regularized curriculum learning scheme that weights samples by lesion size and loss rank.
Key results
- On ISIC2018 small lesions, S³-Mamba reaches 77.13% mIoU versus 64.04% for the VmUnet baseline.
- On the in-house Lymph dataset, S³-Mamba achieves 61.19% mIoU and 75.93% DSC, versus 52.27% and 68.66% for VmUnet.
- Ablation on Lymph shows each component adds gains, with TCMA+CL reaching 59.62% mIoU before the full model's 61.19%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S³-Mamba instead of the strongest listed baseline | ISIC2018 small-lesion segmentation · ISIC2018 small lesions | mIoU: 71.36 -> 77.13 | Helped |
| Use S³-Mamba instead of VmUnet | ISIC2018 large-lesion segmentation · ISIC2018 large lesions | mIoU: 84.67 -> 83.28 | Hurt |
| Use S³-Mamba instead of UNet | CVC-ClinicDB segmentation · CVC-ClinicDB small lesions | mIoU: 69.17 -> 75.40 | Helped |
| Use S³-Mamba instead of MedFormer | Lymph segmentation · Lymph | mIoU: 55.25 -> 61.19 | Helped |
| Add EnCFBlock (ablation) | Baseline · Lymph | mIoU: 52.27 -> 55.70 | Helped |
| Add TCMA (ablation) | Baseline · Lymph | mIoU: 52.27 -> 55.66 | Helped |
| Add curriculum learning (ablation) | Baseline · Lymph | mIoU: 52.27 -> 55.15 | Helped |
| Add TCMA and EnCFBlock (ablation) | Baseline · Lymph | mIoU: 52.27 -> 57.14 | Helped |
| Add TCMA and curriculum learning (ablation) | Baseline · Lymph | mIoU: 52.27 -> 59.62 | Helped |
| Add EnCFBlock, TCMA, and curriculum learning (ablation) | Baseline · Lymph | mIoU: 52.27 -> 61.19 | Helped |
| Add EnCFBlock (ablation) | Baseline · ISIC2018 small lesions | mIoU: 64.04 -> 68.51 | Helped |
| Add TCMA (ablation) | Baseline · ISIC2018 small lesions | mIoU: 64.04 -> 67.97 | Helped |
Surprising findings
- Curriculum learning alone slightly lowered Lymph mIoU (55.15% vs 55.66% for TCMA alone and 55.70% for EnCF alone), though it helped in combinations.
- On the large-lesion ISIC2018 split, S³-Mamba's mIoU (83.28%) is below the VmUnet baseline (84.67%), and its sensitivity is not the highest on several settings.
Limitations
- Small-lesion gains are not uniform: large-lesion mIoU trails VmUnet on ISIC2018, and sensitivity is not best on CVC-ClinicDB and ISIC2018.
- The Lymph dataset is in-house and not released, limiting external reproducibility.
- Curriculum learning alone gave inconsistent gains in the ablations.
Experimental setup
Datasets: ISIC2018 CVC-ClinicDB Lymph (in-house)
Benchmarks: ISIC2018 CVC-ClinicDB Lymph
Metrics: mIoU DSC ACC SPE SEN
Baselines: UNet UNETR++ MedT H2Former ViTSeg SwinUNet SwinUNETR MedFormer VmUnet
Builds on: VMamba VM-UNet curriculum learning
Closest papers by meaning
- Hybrid Mamba for Few-Shot Segmentation (NeurIPS 2024 · similarity 0.85)
- Category Prompt Mamba Network for Nuclei Segmentation and Classification (AAAI 2025 · similarity 0.85)
- Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation (NeurIPS 2025 · similarity 0.85)
- Unified Medical Lesion Segmentation via Self-referring Indicator (CVPR 2025 · similarity 0.83)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.83)
- Multi-view Evidential Learning-based Medical Image Segmentation (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.