Research map / Architectures and efficiency / CNN design

EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation

CVPR 2025 · Architectures and efficiency › CNN design · approach: Dense matching and estimation CNNs · 9 citations (OpenAlex)

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 EffiDec3D, a lightweight 3D segmentation decoder that uniformly reduces channels to the minimum encoder channel count and stops upsampling at half resolution, cutting parameters and FLOPs by over 80% with comparable Dice across 12 medical tasks.

Problem

3D medical segmentation networks such as SwinUNETR, SwinUNETRv2, and 3D UX-Net have high #Params and #FLOPs, largely due to decoder designs with high-resolution layers and large channel counts, limiting real-time and resource-constrained use.

Approach

Replace the heavy decoder with a uniform channel count (the minimum encoder stage channel count, 48 in 3D UX-Net) using residual conv blocks, and limit upsampling to half input resolution followed by a 1x1x1 prediction head and trilinear upsampling, so that the costly full-resolution stages are removed while Dice is largely preserved.

Key results

• On 3D UX-Net, EffiDec3D reduces decoder params by 96.4% (51.7M to 1.8M) and FLOPs by 93.0% (624.2 to 43.7 GFLOPs) relative to the original decoder.
• On FeTA 2021, 3D UX-Net with EffiDec3D reaches 87.97% average DICE vs 87.28% for the original with about 94% fewer parameters.
• On BTCV, 3D UX-Net with EffiDec3D averages 79.25% DICE vs 79.74% original; on MSD, SwinUNETRv2 with EffiDec3D reaches 74.71% vs 73.83% original.

Measured effects reported in the paper

Surprising findings

• Reducing output resolution to half input resolution slightly improved FeTA DICE for some configurations (0.44% gain) while cutting hundreds of GFLOPs.
• Fewer channels (e.g., 48) matched or nearly matched the original decoder's DICE despite far fewer parameters.

Limitations

• Small structures such as esoph agus, veins, and colon show slight Dice drops when resolution is reduced.
• Lower output resolutions (D/8, D/16) degrade performance substantially, so the resolution cut is task-dependent.

Experimental setup

Datasets: FeTA 2021 BTCV 13-organ MSD 10 tasks

Benchmarks: FeTA 2021 BTCV MSD

Metrics: DICE HD95 MParams GFLOPs

Baselines: 3D UX-Net SwinUNETR SwinUNETRv2 MedNeXt M K3 nnUNet nnFormer TransBTS UNETR++ SegFormer3D SlimUNETR UNETR

Builds on: 3D UX-Net SwinUNETR SwinUNETRv2 UNETR

Closest papers by meaning

• E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.88)
• Upping the Game: How 2D U-Net Skip Connections Flip 3D Segmentation (NeurIPS 2024 · similarity 0.88)
• EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation (CVPR 2024 · similarity 0.88)
• Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.87)
• 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)