Research map / Architectures and efficiency / RNNs and normalization
SuperLightNet: Lightweight Parameter Aggregation Network for Multimodal Brain Tumor 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 SuperLightNet, a ~2.97M-parameter encoder-decoder for multimodal MRI brain tumor segmentation using a random multiview drop (RMD) encoder and a learnable residual skip (LRS) decoder, reporting competitive DSC with large reductions in parameters and compute versus SOTA on BraTS2019/2021.
Problem
Multimodal 3D brain tumor segmentation requires heavy 3D convolution, Transformer, or Mamba computation, making it hard to deploy on limited hardware; the goal is to keep accuracy while minimizing parameters and computation.
Approach
A random multiview drop encoder that randomly assigns coronal/axial/sagittal axis Hadamard-product attention branches per step and a decoder replacing transposed convolutions with linear upsampling, grouped MLPs, and learnable group skip and scale residual weights, with channel-split dropRes to cut parameters.
Key results
- On BraTS2019 5-CV, SuperLightNet reaches 82.30% DSC and 73.23% SDC with 2.97M parameters, versus SegMamba's 82.04% DSC and 72.95% SDC at 67.42M.
- On BraTS2021 test (TVE), SuperLightNet achieves 84.05% DSC and 82.99% SDC, a 0.06% DSC gain over SwinUNETR per the paper.
- Computational cost at 4x128^3 input is 0.282 TFlops fwd+bwd versus 0.522 for LightM-Unet, a 45.97% reduction; parameters are reported as 95.59% fewer than SegMamba.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SegMamba with SuperLightNet | Brain tumor segmentation · BraTS2019 validation, 5-CV | DSC: 82.04 -> 82.30 | Helped |
| replace SegMamba with SuperLightNet | Brain tumor segmentation · BraTS2019 validation, 5-CV | SDC: 72.95 -> 73.23 | Helped |
| replace nnUnet with SuperLightNet | Brain tumor segmentation · BraTS2021 Eval | DSC: 87.33 -> 87.45 | Helped |
| replace nnUnet with SuperLightNet | Brain tumor segmentation · BraTS2021 Eval | SDC: 83.21 -> 82.99 | Hurt |
| add bidirectional scanning to THPA (ablation) | THPA + UNETR · BraTS2019 ablation | DSC: 78.84 -> 77.58 | Hurt |
| use three paths with reverse scanning for each instead of bidirectional six-path routing (ablation) | THPA + UNETR · BraTS2019 ablation | DSC: 77.58 -> 76.83 | Hurt |
| reduce hidden dimension from 768 to 384 (ablation) | THPAFR2 + UNETR · BraTS2019 ablation | DSC: 76.83 -> 78.63 | Helped |
| add drop residual, retaining half the channels without feature extraction (ablation) | THPAFR2 + UNETR, 384 dimensions · BraTS2019 ablation | DSC: 78.63 -> 79.02 | Helped |
| replace UNETR decoder with LRSUp (ablation) | THPAFR3 · BraTS2019 ablation | DSC: 79.02 -> 78.39 | Hurt |
| replace linear upsampling with deconvolution (ablation) | THPAFR3 + LRSUp · BraTS2019 ablation | DSC: 78.39 -> 78.68 | Helped |
| use THPAFR3 encoder blocks as decoder blocks (ablation) | THPAFR3 + LRSUp · BraTS2019 ablation | DSC: 78.39 -> 78.74 | Helped |
| use C/12 grouped MLP and grouped weighting for skip and residual connections (ablation) | THPAFR3 + LRSUp · BraTS2019 ablation | DSC: 78.39 -> 79.05 | Helped |
Surprising findings
- Reported parameter reduction vs. SegMamba (95.59%) and vs. LightM-Unet (40.83%) are inconsistent with the 2.97M vs 67.42M and 5.02M figures in the tables for some comparisons, and the paper cites multiple different percentage figures for the same comparisons.
Limitations
- Accuracy gains over SOTA are marginal (about 0.21% average), and the paper itself notes the need for more accurate lightweight networks.
- Evaluation on only two BraTS datasets; the BraTS2021 results use a custom 7:1:2 split and BraTS2019 uses validation folds rather than a held-out test set.
- Ablation design choices are tuned on the reported validation results, raising concerns about selection bias.
Experimental setup
Datasets: BraTS2019 BraTS2021
Benchmarks: BraTS2019 BraTS2021
Metrics: DSC SDC parameter count TFlops MACs training memory inference memory training time
Baselines: 3D UNet TransBTS nnUNet SwinUNETR TransUNet MedNeXt-M SegMamba LightM-Unet
Builds on: UNETR U-Mamba MedNeXt KAN
Closest papers by meaning
- Incomplete Multi-modal Brain Tumor Segmentation via Learnable Sorting State Space Model (CVPR 2025 · similarity 0.84)
- E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.83)
- Revisiting MAE Pre-training for 3D Medical Image Segmentation (CVPR 2025 · similarity 0.83)
- Federated Modality-Specific Encoders and Multimodal Anchors for Personalized Brain Tumor Segmentation (AAAI 2024 · similarity 0.83)
- Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.83)
- One-to-Multiple: A Progressive Style Transfer Unsupervised Domain-Adaptive Framework for Kidney Tumor Segmentation (NeurIPS 2024 · 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.