Research map / Detection and segmentation / CNN features
Rethinking Decoder Design: Improving Biomarker Segmentation Using Depth-to-Space Restoration and Residual Linear Attention
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 MCADS, a U-Net-style decoder with depth-to-space upsampling, residual linear attention and channel-spatial attention, paired with a modified U2-Net encoder, for biomarker segmentation trained from scratch.
Problem
Biomarker segmentation in medical images (nuclei, glands, mitochondria) where end-to-end models underperform pre-trained encoders, due to poor encoder-to-decoder feature transfer and inefficient decoders.
Approach
A decoder that uses depth-to-space upsampling in distal stages to avoid information loss, residual linear attention blocks to refine skip connections, and a channel-and-spatial attention block, combined with a U2-Net encoder augmented with attention; trained end-to-end from scratch.
Key results
- On MoNuSeg, the method reaches 74.04 IoU and 85.04 Dice, versus 71.28 IoU for EMCAD (reproduced).
- On Electron Microscopy, the method reaches 88.89 IoU versus 86.31 for the best reproduced baseline (U2-Net).
- On zero-shot TNBC, the method reaches 69.16 IoU, below UN-SAM (72.27) which was fine-tuned on TNBC, but above EMCAD (65.13).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed method instead of SAC 0-expert | Biomarker segmentation · MoNuSeg | IoU: 72.61 -> 74.04 | Helped |
| Use the proposed method instead of SAC 0-expert | Biomarker segmentation · DSB | IoU: 87.32 -> 89.48 | Helped |
| Use the proposed method instead of UNet++ | Biomarker segmentation · Electron Microscopy | IoU: 87.01 -> 88.89 | Helped |
| Use the proposed method instead of UN-SAM | Biomarker segmentation · TNBC (zero-shot) | IoU: 72.27 -> 69.16 | Hurt |
| Add Inner Attention to the baseline U2-Net encoder (ablation) | U2-Net encoder · MoNuSeg | IoU: 72.03 -> 72.81 | Helped |
| Add CASAB to the encoder with Inner Attention (ablation) | U2-Net encoder · MoNuSeg | IoU: 72.81 -> 74.04 | Helped |
| Replace transpose convolutions at all decoder stages with EUB (ablation) | MCADS decoder · MoNuSeg | IoU: 70.52 -> 72.23 | Helped |
| Add DSUB at the bridge while keeping EUB at the remaining stages (ablation) | MCADS decoder · MoNuSeg | IoU: 72.23 -> 72.96 | Helped |
| Add DSUB at stage 4 (ablation) | MCADS decoder · MoNuSeg | IoU: 72.96 -> 74.04 | Helped |
| Add a third DSUB stage (ablation) | MCADS decoder · MoNuSeg | IoU: 74.04 -> 74.10 | Helped |
| Add a fourth DSUB stage (ablation) | MCADS decoder · MoNuSeg | IoU: 74.10 -> 74.79 | Helped |
| Add a fifth DSUB stage (ablation) | MCADS decoder · MoNuSeg | IoU: 74.79 -> 74.88 | Helped |
Surprising findings
- Pre-trained encoder based SOTA methods such as EMCAD and PVT-CASCADE underperformed the proposed end-to-end model trained from scratch.
- Adding more DSUB stages beyond the first two gave minimal accuracy gains while greatly increasing compute.
Limitations
- TNBC results are zero-shot, and the authors note that training on TNBC could improve accuracy but was not done.
- Ablation experiments used limited hardware, so deeper DSUB configurations were not fully explored.
- SAM-based comparison numbers were copied from original papers for some methods, not reproduced under the same setup.
Experimental setup
Datasets: MoNuSeg DSB 2018 Electron Microscopy TNBC
Benchmarks: MoNuSeg DSB 2018 Electron Microscopy TNBC
Metrics: IoU Dice Precision Recall FOR HD95 ASD
Baselines: U-Net Hover-Net UNet++ U2-Net nnU-Net FusionU-Net SegResNet UCTransNet TransUNet Swin-Unet PVT-CASCADE EMCAD
Builds on: U2-Net EMCAD U-Net CBAM Depth-to-Space
Closest papers by meaning
- EMCAD: Efficient Multi-scale Convolutional Attention Decoding for Medical Image Segmentation (CVPR 2024 · similarity 0.86)
- nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark (CVPR 2025 · similarity 0.85)
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.85)
- ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement (AAAI 2025 · similarity 0.84)
- A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.84)
- MEATRD: Multimodal Anomalous Tissue Region Detection Enhanced with Spatial Transcriptomics (AAAI 2025 · similarity 0.84)
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.