Research map / Architectures and efficiency / CNN design
E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image 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 E2ENet, a 3D medical segmentation network that fuses multi-scale features with dynamic sparse training (DSFF) and uses a restricted depth-shift in 1x3x3 convolutions to get accuracy comparable to nnUNet with ~3x fewer parameters and lower FLOPs.
Problem
Accurate 3D medical image segmentation with deep networks is costly in parameters and FLOPs, hindering deployment on resource-limited hardware; existing multi-scale fusion (UNet++, NAS-based DiNTS) is redundant or expensive to search.
Approach
Fuse multi-scale features in a grid of fusion nodes whose channel connections are a binary sparse mask trained with dynamic sparse training (prune low-L1 kernels, randomly regrow the same number each period), combined with a depth-only channel shift so that cheap 1x3x3 convolutions capture inter-slice information.
Key results
- On AMOS-CT validation, E2ENet (s=0.8) reaches 90.0 mDice with 9.44M params and 778.74G FLOPs vs nnUNet 90.0 mDice, 30.76M params, 1067.89G FLOPs.
- On BraTS 5-fold CV, E2ENet (S=0.9) reaches 74.3 mDice with 7.63M params vs nnUNet 74.1 mDice with 31.20M params.
- The abstract reports over 69% parameter saving and 27% FLOPs saving compared with the previous best-performing method on AMOS-CT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase feature sparsity from S=0.7 to S=0.9 (ablation) | E2ENet* · AMOS-CT validation | mDice: 90.3 -> 89.9 | Hurt |
| increase feature sparsity from S=0.7 to S=0.9 (ablation) | E2ENet · BraTS MSD 5-fold training set | mDice: 74.5 -> 74.3 | No clear effect |
| add DSFF (ablation) | E2ENet with restricted depth-shift · AMOS-CT validation | mDice: 90.2 -> 90.1 | No clear effect |
| add DSFF (ablation) | E2ENet with restricted depth-shift · BraTS MSD 5-fold training set | mDice: 74.1 -> 74.5 | Helped |
| add restricted depth-shift (ablation) | E2ENet without DSFF · AMOS-CT validation | mDice: 88.6 -> 90.2 | Helped |
| add restricted depth-shift (ablation) | E2ENet with DSFF · AMOS-CT validation | mDice: 88.2 -> 90.1 | Helped |
| increase shift size to (-7, 0, 7) (ablation) | E2ENet · AMOS-CT validation | mDice: 90.1 -> 87.6 | Hurt |
| set DSFF update interval to 600 or 1800 (ablation) | E2ENet · AMOS-CT validation | mDice: 90.1 -> 90.0; 90.1 -> 89.9 | No clear effect |
| add restricted depth-shift (ablation) | E2ENet with DSFF · BraTS MSD 5-fold training set | mDice: 73.9 -> 74.5 | Helped |
| increase level-1 channels from 48 to 58 (ablation) | E2ENet · AMOS-CT capacity analysis | mDice: 90.0 -> 90.4 | Helped |
| reduce level-1 channels from 32 to 27 and feature scales from 5 to 4 (ablation) | nnUNet · AMOS-CT capacity analysis | mDice: 90.0 -> 89.7 | Hurt |
Surprising findings
- Removing the restricted depth-shift and using full 3x3x3 kernels gives similar mDice to 1x3x3 with depth shift on AMOS-CT, but the shifted variant is much cheaper.
- E2ENet with static (non-evolving) sparse topology reports identical mDice to the dynamic version on AMOS-CT validation at s=0.9.
Limitations
- Sparsity is implemented with binary masks on dense weights, so no real training speedup is achieved.
- Efficiency gains (FLOPs) are analytical estimates; GPU speedups are not demonstrated.
- Gains over nnUNet on accuracy are marginal; mNSD is slightly lower than nnUNet in several settings.
Experimental setup
Datasets: AMOS-CT BraTS (MSD) BTCV AMOS-MRI
Benchmarks: AMOS-CT BraTS (MSD) BTCV AMOS-MRI
Metrics: mDice mNSD Params FLOPs PT score
Baselines: nnUNet CoTr nnFormer UNETR Swin UNETR VNet UNet++ DiNTS DeepLabv3 MedFormer
Builds on: nnUNet UNet++ temporal shift dynamic sparse training DiNTS
Closest papers by meaning
- EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- FuseUNet: A Multi-Scale Feature Fusion Method for U-like Networks (ICML 2025 · similarity 0.87)
- DDViT: Double-Level Fusion Domain Adapter Vision Transformer (Student Abstract) (AAAI 2024 · similarity 0.86)
- Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.86)
- VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging (CVPR 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)
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.