Research map / Detection and segmentation / CNN features
FuseUNet: A Multi-Scale Feature Fusion Method for U-like Networks
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 FuseUNet, which treats U-Net skip connections as nodes of an initial value problem and fuses multi-scale features via an adaptive linear multistep (Adams-Bashforth/Moulton with predictor-corrector) discretization of nmODEs, reducing parameters while maintaining segmentation performance.
Problem
Skip connections in U-Net-family networks only fuse same-scale features and rely on simple concatenation or addition, limiting multi-scale interaction and feature utilization in medical image segmentation.
Approach
Reinterpret the U-Net decoding as an initial value problem over skip-connection nodes and apply adaptive high-order linear multistep methods (Adams-Bashforth/Moulton predictor-corrector) with nmODE blocks, so each decoder stage fuses features from multiple previous scales.
Key results
- On ACDC with nn-UNet backbone, FuseUNet reaches 91.57% average Dice vs 91.54% for nn-UNet while cutting parameters from 31.2M to 14.0M (54.9% reduction).
- On KiTS, FuseUNet achieves 86.19% average Dice vs 86.04% for nn-UNet; on MSD with UNETR backbone it reaches 72.6% vs 71.1%.
- On ISIC2018 with UltraLight VM-UNet backbone, FuseUNet reports 89.77% Dice and reduces parameters by 29.8% (0.049M to 0.036M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FuseUNet multi-scale feature fusion to nn-UNet | nn-UNet · ACDC | Dice avg: 91.54 -> 91.57 | No clear effect |
| add FuseUNet multi-scale feature fusion to nn-UNet | nn-UNet · KiTS | Dice avg: 86.04 -> 86.19 | No clear effect |
| add FuseUNet multi-scale feature fusion to UNETR | UNETR · MSD brain tumor | Dice avg: 71.1 -> 72.6 | Helped |
| add FuseUNet multi-scale feature fusion to UltraLight VM-UNet | UltraLight VM-UNet · ISIC2017 | Dice: 90.64 -> 90.69 | No clear effect |
| add FuseUNet multi-scale feature fusion to UltraLight VM-UNet | UltraLight VM-UNet · ISIC2018 | Dice: 89.40 -> 89.77 | Helped |
| increase maximum feature-fusion order from 1 to 4 (ablation) | FuseUNet · KiTS | Dice: 84.8 -> 86.7 | Helped |
| increase maximum feature-fusion order from 1 to 4 (ablation) | FuseUNet · ACDC | Dice: 91.75 -> 92.05 | Helped |
| increase maximum feature-fusion order from 1 to 4 (ablation) | FuseUNet · MSD brain tumor | Dice: 71.22 -> 71.75 | Helped |
| increase maximum feature-fusion order from 1 to 4 (ablation) | FuseUNet · ISIC2017 | Dice: 89.25 -> 90.69 | Helped |
| increase maximum feature-fusion order from 1 to 4 (ablation) | FuseUNet · ISIC2018 | Dice: 88.63 -> 89.78 | Helped |
| increase memory channels from N to 2N (ablation) | FuseUNet · KiTS first fold | Dice: 85.86 -> 86.74 | Helped |
| increase memory channels from 2N to 3N (ablation) | FuseUNet · KiTS first fold | Dice: 86.74 -> 86.63 | Hurt |
Surprising findings
- FLOPs values in the original paper were under-estimated by a factor of 10 (post-publication correction), and FuseUNet slightly increases GFLOPs on the lightweight 2D backbone due to interpolation.
- On ACDC and KiTS, paired t-tests show no statistically significant difference between FuseUNet and nn-UNet (p=0.80 and p=0.58).
Limitations
- Must store multiple historical solutions, causing high memory consumption, especially with many target classes.
- Parameter reduction is modest for Transformer backbones since the method only modifies skip connections, not the encoder.
- Slight GFLOPs increase on lightweight networks due to interpolation for skip-connection alignment.
Experimental setup
Datasets: ACDC KiTS2023 MSD brain tumor ISIC2017 ISIC2018
Benchmarks: ACDC KiTS2023 MSD brain tumor ISIC2017 ISIC2018
Metrics: Dice sensitivity specificity accuracy parameters (M) GFLOPs
Baselines: nn-UNet UNETR CoTr Swin-UNETR U-Mamba STU-Net-L nnFormer UNet UNet3+ TransUNet TransFuse MALUNet
Builds on: U-Net nmODEs Adams-Bashforth/Adams-Moulton linear multistep method nnU-Net UNETR UltraLight VM-UNet
Closest papers by meaning
- Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation (AAAI 2025 · similarity 0.88)
- Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.87)
- E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.87)
- LoMix: Learnable Weighted Multi-Scale Logits Mixing for Medical Image Segmentation (NeurIPS 2025 · similarity 0.86)
- nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark (CVPR 2025 · similarity 0.86)
- U-KAN Makes Strong Backbone for Medical Image Segmentation and Generation (AAAI 2025 · similarity 0.85)
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.