Research map / Architectures and efficiency / CNN design
EMCAD: Efficient Multi-scale Convolutional Attention Decoding for 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 EMCAD, an efficient multi-scale convolutional attention decoder for 2D medical image segmentation that pairs with PVTv2 encoders and cuts parameters and FLOPs substantially versus CASCADE.
Problem
Decoders for medical image segmentation achieve strong accuracy but are computationally expensive, limiting use in resource-constrained settings.
Approach
A cascaded decoder built from depth-wise multi-scale convolution blocks (MSCAM with channel, spatial, and multi-scale depth-wise convolution), large-kernel grouped attention gates replacing point-wise gating, and efficient depth-wise up-convolution blocks, trained with a multi-stage combinatorial loss.
Key results
- PVT-EMCAD-B2 reaches 91.10% average DICE over 10 binary segmentation datasets with 26.76M params and 5.6G FLOPs.
- On Synapse, EMCAD with PVTv2-B2 gives 83.63% DICE using 1.91M decoder params and 0.381G FLOPs, versus CASCADE's 82.78% with 9.27M params, a 79.4% parameter and 80.3% FLOPs reduction.
- PVT-EMCAD-B2 achieves 92.12% average DICE on ACDC, slightly above Cascaded MERIT (91.85%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the CASCADE decoder with EMCAD | PVTv2-B2 · Synapse Multi-organ | DICE: 82.78 -> 83.63 | Helped |
| replace the CASCADE decoder with EMCAD | PVTv2-B0 · Synapse Multi-organ | DICE: 80.54 -> 81.97 | Helped |
| add a cascaded decoder (ablation) | PVTv2-B2 encoder · Synapse Multi-organ | DICE: 80.10±0.2 -> 81.08±0.2 | Helped |
| add LGAG to the cascaded decoder (ablation) | PVTv2-B2 encoder · Synapse Multi-organ | DICE: 81.08±0.2 -> 81.92±0.2 | Helped |
| add MSCAM to the cascaded decoder (ablation) | PVTv2-B2 encoder · Synapse Multi-organ | DICE: 81.08±0.2 -> 82.86±0.3 | Helped |
| add LGAG to the cascaded decoder with MSCAM (ablation) | PVTv2-B2 encoder · Synapse Multi-organ | DICE: 82.86±0.3 -> 83.63±0.3 | Helped |
| use kernels [1, 3] instead of [1] (ablation) | MSDC in EMCAD · Synapse Multi-organ | DICE: 82.43 -> 82.98 | Helped |
| use kernels [1, 3, 5] instead of [1, 3] (ablation) | MSDC in EMCAD · Synapse Multi-organ | DICE: 82.98 -> 83.63 | Helped |
| use kernels [1, 3, 5] instead of [1, 3] (ablation) | MSDC in EMCAD · ClinicDB | DICE: 95.13 -> 95.21 | Helped |
| add larger kernels [7, 9] to [1, 3, 5] (ablation) | MSDC in EMCAD · Synapse Multi-organ and ClinicDB | DICE: 83.63 -> 83.34; 95.21 -> 95.07 | Hurt |
| use two 3×3 kernels instead of kernels [1, 3] (ablation) | MSDC in EMCAD · Synapse Multi-organ | DICE: 82.98 -> 82.81 | Hurt |
Limitations
- Evaluated only on 2D medical image segmentation tasks.
- Gains over strongest baselines on some datasets are small or absent (e.g., PVT-CASCADE better on some polyp datasets).
Experimental setup
Datasets: ClinicDB Kvasir ColonDB ETIS BKAI ISIC17 ISIC18 BUSI EM DSB18 Synapse multi-organ ACDC
Benchmarks: ClinicDB Kvasir ColonDB ETIS BKAI ISIC17 ISIC18 BUSI EM DSB18 Synapse multi-organ ACDC
Metrics: DICE HD95 mIoU #Params #FLOPs
Baselines: UNet UNet++ AttnUNet DeepLabv3+ PraNet CaraNet UACANet-L SSFormer-L PolypPVT TransUNet SwinUNet TransFuse
Builds on: CASCADE PVTv2 MERIT Attention U-Net MobileNetV2
Closest papers by meaning
- Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline (CVPR 2025 · similarity 0.88)
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.88)
- EffiDec3D: An Optimized Decoder for High-Performance and Efficient 3D Medical Image Segmentation (CVPR 2025 · similarity 0.88)
- Rethinking Decoder Design: Improving Biomarker Segmentation Using Depth-to-Space Restoration and Residual Linear Attention (CVPR 2025 · similarity 0.86)
- Generative Medical Segmentation (AAAI 2025 · similarity 0.85)
- nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark (CVPR 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.