MedCAGD: Context-Aware Gated Decoder for Efficient Medical Image Segmentation

Conference: ECCV 2026
arXiv: 2607.00409
Code: https://github.com/saadwazir/MedCAGD
Area: Medical Image
Keywords: Medical Image Segmentation, Decoder Design, Gated Skip Connections, Context Aggregation, Channel Attention

TL;DR

MedCAGD proposes a decoder-centric medical image segmentation framework that systematically regulates skip connection fusion and multi-scale context aggregation through a context-aware gated decoder. At a lightweight computational cost of 30.60M parameters and 5.0 GFLOPs, it consistently outperforms CNN, Transformer, Mamba, and SOTA decoder methods across 11 medical image segmentation benchmarks.

Background & Motivation

The dominant paradigm for medical image segmentation is the encoder-decoder architecture (represented by U-Net). In recent years, research focus has heavily biased toward the encoder side—large-scale pre-training, Transformer encoders, and Mamba state space models have continuously improved feature extraction quality. However, the segmentation accuracy encounters a significant bottleneck at the decoder side: issues such as inaccurate multi-scale feature alignment, difficulty in effectively injecting global context, and insufficient boundary refinement mean that even when the encoder extracts high-quality features, the final pixel-level predictions still suffer from semantic misalignment and blurred boundaries.

Key Challenge: While encoder capability has become sufficiently powerful (pre-trained models like PVTv2, ConvNeXt, and MaxViT provide rich multi-scale features), decoder design remains stuck in naive upsampling coupled with direct skip connection concatenation. This lacks systematic regulation over feature fusion and context aggregation. Prior decoder improvements (such as EMCAD and MCADS) attempt to enhance feature fusion via attention mechanisms, but they either achieve limited gains at a large computational cost, or rely on additive fusion that lacks structured handling of the semantic gap between encoder and decoder—simply adding features from both sides and passing them through a sigmoid essentially assumes that encoder-decoder features are directly additive semantically, an assumption that does not hold at shallower layers.

Core Idea: Regard the decoder design as the primary controlling factor for segmentation accuracy, and propose a structured context-aware gated decoder. Through gated skip connections (modeling multiplicative consistency combined with multi-scale spatial competition), multi-level context aggregation injection, and refinement modules with channel recalibration, it allows strong encoder features to be more faithfully converted into spatially consistent pixel predictions without relying on larger encoders or task-specific fine-tuning.

Method

Overall Architecture

MedCAGD is an encoder-decoder architecture, but its innovation is entirely concentrated on the decoder side. The encoder (defaulting to PVTv2-B2, pre-trained on ImageNet) extracts hierarchical features \(\{c_i\}_{i=1}^4\) at four scales, which are projected to unified decoder channel dimensions (64, 128, 320, 512) via a Universal Feature Projection (\(1\times1\) convolution). Decoding starts from the deepest feature, initialized and refined through a Bottleneck, then progressively upsampled, selectively fused with encoder skip features via the SCA-Gate, injected with multi-level context aggregation information, and finally refined via the Refinement Block. This sequence (upsampling + gated skip fusion + context aggregation + refinement) is executed repeatedly across three decoding stages. The output of the final stage is fed into the segmentation head to yield the main prediction, while intermediate stages connect auxiliary segmentation heads and edge prediction heads to implement deep supervision. During inference, only the final segmentation prediction is retained.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Image 224×224"] --> B["PVTv2-B2 Encoder<br/>4-Scale Hierarchical Features"]
    B --> C["Universal Feature Projection<br/>1×1 Conv to Unify Channel Dimensions"]
    C --> D["Bottleneck<br/>ECA-MSP + Conv + RA"]
    C --> E["Multi-Scale Skip Features p1-p4"]
    D --> F["Upsampling 2×"]
    E --> G["SCA-Gate<br/>Gated Selective Skip Fusion"]
    F --> G
    C --> H["Context Aggregator<br/>Multi-Level Context Aggregation + RA"]
    H --> I["Residual Injection of Global Context"]
    G --> I
    I --> J["Refinement Block<br/>DWConv + PWConv + ECA-MSP"]
    J -->|"Repeat 3 Decoding Stages"| F
    J --> K["Segmentation Head: Main Prediction"]
    J --> L["Auxiliary Segmentation Head<br/>Deep Supervision"]
    J --> M["Edge Prediction Head<br/>Edge Supervision"]

Key Designs

1. Bottleneck with Global Context Injection: Inject global semantics at the starting point of decoding to prevent info collapse of the deepest features

The encoder's deepest feature \(F_4\) has the largest receptive field but the lowest spatial resolution (e.g., \(7\times7\)). If upsampling is directly performed to start decoding, local details are almost completely lost, and a validation of its own optical semantic quality is lacking. The Bottleneck processes \(F_4\) in three steps at the start of decoding: first, ECA-MSP (efficient channel attention with multi-scale pooling) is used to recalibrate channel responses, assigning higher weights to semantic channels crucial for the segmentation task; next, a standard convolution is utilized for local refinement; finally, global context is injected via Residual Attention (RA). The overall transformation is \(B = \mathcal{R}(\rho(\mathcal{E}(F_4)))\), where \(\mathcal{E}\) represents ECA-MSP, \(\rho\) represents convolutional refinement, and \(\mathcal{R}\) represents RA. Ablation studies show that adding the BT alone improves the Synapse Dice score from 73.91 to 75.53 (+1.62), acting as the first anchor of model performance.

The key improvement of ECA-MSP lies in multi-scale pooling: unlike standard SE-Net, which only uses global average pooling to obtain a single scalar descriptor, ECA-MSP extracts channel statistics using three pooling scales \(\mathcal{S}=\{1,2,4\}\). The \(1\times1\) pooling yields a global semantic descriptor, while \(2\times2\) and \(4\times4\) poolings capture coarse-grained local context clues. The three-way descriptors individually model local cross-channel interactions through 1D convolutions (without dimensionality reduction, avoiding information bottlenecks), and their average is passed through a sigmoid to generate attention weights. This multi-granularity design enables channel recalibration to perceive both global and local contexts simultaneously, being more robust than single-scale SE/ECA, which might erroneously apply the same channel weight across all spatial positions because global pooling flattens out local variations.

2. SCA-Gate (Spatially Competitive Attention Gate): Replace additive gating with multiplicative consistency and spatial competition to selectively regulate skip connections

Traditional skip connections (such as direct concatenation in U-Net) implicitly assume that encoder and decoder features are semantically compatible. In reality, a significant semantic gap exists between them—shallow encoder features are rich in texture and edges but semantically weak, while deep decoder features are semantically strong but spatially blurry. Simple concatenation introduces semantic noise into the decoding process. SCA-Gate models skip connections as a learnable, selective feature regulation mechanism rather than a passive pathway.

Specific formulation: Let the decoder feature be \(g\), and its corresponding encoder skip feature be \(x\). Both are first subjected to channel recalibration using ECA-MSP, and then projected into a shared latent space. Their interaction is modeled as element-wise multiplication \(f = \theta(\mathcal{E}(g)) \odot \phi(\mathcal{E}(x))\)—multiplication instead of addition, because multiplication naturally requires both sides to be active simultaneously at the same spatial position to generate a strong response, which is equivalent to "a signal agreed upon by both the encoder and decoder". The gating signal \(\mathcal{H}(f, g, x)\) is formed by the continuous multiplication of three terms:

\[\mathcal{H}(f, g, x) = f \odot (1 + \mathcal{G}(g, x)) \odot (1 + \mathcal{S}(f))\]

where \(\mathcal{G}\) is global channel modulation (extracting global channel statistics from the joint encoder-decoder representation), and \(\mathcal{S}\) is multi-scale spatial competition—aggregating multi-scale neighborhood responses using parallel depthwise separable convolutions \(\mathcal{D}_3(f)\) and \(\mathcal{D}_5(f)\) with kernels 3 and 5, followed by temperature-scaled softmax normalization to form spatial-dimensional competition: within the multi-scale neighborhood, only truly important spatial locations are activated, while background and noise regions are suppressed. The final attention mask \(\sigma(\mathcal{H})\) is multiplied back to the skip feature \(x\) to achieve selective transmission.

Compared to the additive fusion and sigmoid mask of Attention U-Net, the triple mechanism of SCA-Gate (multiplicative consistency + global modulation + spatial competition) provides more structured feature filtering. Table 5 shows that SCA-Gate achieves 87.00 Dice on Synapse, outperforming the Attention U-Net Gate (85.20) by 1.8 points and EMCAD's LGAG (84.51) by 2.49 points, while requiring fewer parameters (30.60M vs 30.94M LGAG).

3. Multi-level Context Aggregator (CA) with Residual Attention: Allow each decoding stage to "see" global semantics

Skip connections only transfer information between corresponding levels (e.g., encoder-stage-2 to decoder-stage-2), but effective decoding also requires cross-level global perception—deep refinement guided by shallow details, and shallow localization clues reinforcing deep semantics. The Context Aggregator aggregates the projected features \(\{F_k\}_{k=1}^K\) from all \(K\) encoder levels: each is first projected to a unified channel dimension via \(1\times1\) pointwise convolution, matching the spatial resolution of the current decoder stage through interpolation. Their average is then globally refined by RA, yielding \(F_{\text{ctx}} = \mathcal{R}(\frac{1}{K}\sum_{k=1}^K \mathcal{P}_k(F_k))\). This global context representation is injected into each decoding stage via residual addition, offering stage-independent global semantic guidance.

RA (Residual Attention) itself is a lightweight non-local attention module: for an input feature map \(X\), it is first projected using a pointwise convolution \(\mathcal{P}_0\) and then normalized via softmax to obtain a spatial importance distribution of dimension \(HW\). A weighted sum is performed to obtain a global context descriptor, which is then mixed across channels via two pointwise convolution layers (with intermediate dimension reduction and a non-linear activation \(\delta\)), and finally added back residually: \(Y = X + \mathcal{P}_2(\delta(\mathcal{P}_1(\sum_{i=1}^{HW} \text{Softmax}(\mathcal{P}_0(X))_i X_i)))\). The residual form preserves local structures while efficiently injecting global information.

CA and SCA-Gate are complementary: SCA-Gate performs "point-to-point" selective fusion of same-layer features, while CA conducts "surface-like" cross-layer global semantic injection. In the ablation study, CA (with RA) alone elevates Synapse Dice from 73.91 to 81.03 (+7.12), and when combined with BT, it further increases to 83.57.

4. Refinement Block (RB): Lightweight local refinement and channel recalibration to stabilize decoder feature propagation

After fusing skip features and global context, each decoding stage goes through a Refinement Block for final local enhancement. The pipeline of the RB is: depthwise separable convolution (spatial filtering, kernel 3×3) \(\rightarrow\) Group Normalization + SiLU \(\rightarrow\) pointwise convolution (\(1\times1\), channel mixing) \(\rightarrow\) Group Normalization + SiLU \(\rightarrow\) ECA-MSP (channel recalibration). The depthwise separable convolution significantly reduces the computation of spatial convolutions, while ECA-MSP performs final channel importance ranking on the refined features. RB serves as a quick check of local consistency and channel quality of the fused features at the end of each decoding stage—ablation shows that adding RB improves Synapse Dice from 83.57 to 85.19 (+1.62).

A Complete Example: Synapse Multi-Organ Abdominal Segmentation

Here is an execution flow for an input \(224\times224\) abdominal CT slice. The PVTv2-B2 encoder outputs 4 scales of features: \(c_1\) (\(56\times56\), 64 channels), \(c_2\) (\(28\times28\), 128 channels), \(c_3\) (\(14\times14\), 320 channels), and \(c_4\) (\(7\times7\), 512 channels). The Universal Feature Projection projects them into unified dimensions \(\{p_i\}_{i=1}^4\).

First, \(p_4\) (\(7\times7\), 512 channels) enters the Bottleneck: ECA-MSP recalibrates channels using three pooling scales \(\{1,2,4\}\)—the large pooling (\(1\times1\)) captures "whether this channel is globally important for organ segmentation", while the small pooling (\(4\times4\)) captures channel preferences for local texture. After recalibration, it is refined by a \(1\times1\) convolution and then experiences global context injection via RA: a softmax computes the attention correlation of \(7\times7=49\) spatial locations, which are weighted and summed to form a global descriptor, which is residually added back after channel mixing. At this stage, the decoder feature \(d_4\) possesses global semantics.

Stage 3 Decoding: \(d_4\) is upsampled to \(14\times14\). The SCA-Gate fuses it with \(p_3\)—multiplicative consistency ensures that only locations co-activated by both encoder and decoder (e.g., liver boundary) yield high gate values, and multi-scale spatial competition further suppresses noisy areas. Concurrently, the CA aggregates all of \(p_1\) to \(p_4\) to \(14\times14\) global context and residually injects it into \(d_3\). Finally, the RB polishes the feature map using depthwise convolution and channel attention. This stage produces auxiliary segmentation and edge predictions.

Stage 2 (\(28\times28\)) and Stage 1 (\(56\times56\)) repeat the identical process. Ultimately, the \(56\times56\) feature is map-projected via the segmentation head to yield 8-class organ predictions (aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, stomach), and upsampled back to \(224\times224\). The entire process initiates from the \(7\times7\) global semantics of the Bottleneck, recovering spatial details stage-by-stage. Each stage simultaneously receives selective local information from the same level's skip connection (SCA-Gate) and global semantic guidance across all levels (CA), undergoing a final local consistency test in the Refinement Block.

Loss & Training

All segmentation and edge predictions are optimized using binary cross-entropy (BCE) loss. The total loss is a weighted combination of three components: the main segmentation loss \(\mathcal{L}_{\text{main}}\) + deep supervision losses \(\sum_{i=1}^3 \mathcal{L}_{\text{aux}}^i\) (using auxiliary segmentation heads from 3 intermediate decoding stages) + edge supervision losses \(\sum_{i=1}^3 \mathcal{L}_{\text{edge}}^i\) (using edge prediction heads from the same 3 stages, with edge ground truth binarized and extracted from the GT mask). The optimizer is AdamW, with a learning rate of \(1\times10^{-4}\) and a batch size of 16. Training runs for 300+ epochs, and the optimal model is selected based on the validation set Dice score. Inputs are uniformly resized to \(224\times224\), and online data augmentations include random rotation, horizontal/vertical flipping, and random cropping. For high-resolution datasets (DRIVE/FIVES fundus vessels), training is conducted with \(256\times256\) overlapping patches (stride 128); for CellSeg, \(384\times384\) overlapping patches (stride 192) are used, reconstructive patch predictions being stitched during inference to generate the complete segmentation map.

Key Experimental Results

Main Results

9 Binary Segmentation Benchmarks (Average Dice, %): The table covers 6 task categories: skin lesion (ISIC17/18), polyp (ETIS/ColonDB), fundus vessels (DRIVE/FIVES), ultrasound neoplasm (BUSI), thyroid ultrasound (ThyroidXL), and microscopic cells (CellSeg).

Method	Params	FLOPs	Skin Mean	Polyp Mean	Fundus Mean	Neoplasm Mean	Cell	9-Set Mean
U-Net	34.53M	65.53G	84.87	80.40	73.49	72.60	71.52	77.14
nnU-Net	31.29M	55.26G	85.88	85.88	75.77	81.27	83.53	82.34
TransUNet	105.32M	38.52G	87.08	89.71	79.26	82.04	79.08	83.92
EMCAD	26.76M	5.60G	88.46	92.30	79.83	81.79	79.13	84.87
MCADS	50.90M	61.89G	87.58	91.81	77.24	83.18	86.68	85.14
MedCAGD	30.60M	5.00G	89.09	93.37	84.57	85.75	86.61	88.01

Synapse Multi-Organ CT Segmentation (8 Classes, Dice/IoU/HD95):

Method	Dice (%)	IoU (%)	HD95	Aorta	Gallbladder	Left Kidney	Right Kidney	Liver	Pancreas	Spleen	Stomach
nnU-Net	75.33	71.47	19.34	77.06	73.27	76.34	84.53	79.98	73.34	77.62	60.52
TransUNet	77.61	67.32	26.90	86.56	60.43	80.54	78.53	94.33	58.47	87.06	75.00
EMCAD	83.63	74.65	15.68	88.14	68.87	88.08	84.10	95.26	68.51	92.17	83.92
MCADS	85.03	81.71	11.11	90.81	86.07	86.77	83.24	87.66	83.55	85.74	76.38
MedCAGD	87.00	83.77	14.39	92.28	90.31	89.72	87.21	91.02	82.08	86.91	76.51

ACDC Cardiac MRI Segmentation (3 Classes, Dice/IoU/HD95):

Method	Dice (%)	IoU (%)	HD95	Right Ventricle	Myocardium	Left Ventricle
nnU-Net	82.66	74.27	6.17	79.00	81.01	87.97
EMCAD	85.07	77.73	5.25	81.58	84.23	89.42
MCADS	84.51	76.92	5.56	81.16	83.27	89.09
UCTransNet	84.89	77.57	5.70	80.94	84.11	89.62
MedCAGD	87.54	80.96	4.41	85.27	86.23	91.11

Ablation Study

Component Ablation (Synapse Dice / CellSeg Dice, %): The baseline is the PVTv2-B2 encoder + naive U-Net-style decoder (only upsampling + direct skip connection concatenation).

BT	CA (w/ RA)	CA (w/o RA)	RB	SCA-Gate	Synapse	CellSeg
					73.91	81.07
Y					75.53	82.80
Y		Y			79.38	82.60
	Y				81.03	82.47
Y	Y				83.57	84.28
Y	Y		Y		85.19	84.62
Y	Y		Y	Y	87.00	86.61

Comparison of skip connection attention mechanisms (Synapse / CellSeg Dice, %): on the baseline already containing BT + CA + RB, different skip attention modules are substituted.

Skip Attention Scheme	Params	FLOPs	Synapse	CellSeg
Attention U-Net Gate	29.54M	4.63G	85.20	83.19
Attn Gate + ECA	29.58M	4.63G	85.38	80.77
Attn Gate + ECA-MSP	29.45M	4.63G	85.46	84.65
LGAG (EMCAD)	30.94M	4.99G	84.51	81.99
RLAB (MCADS)	30.96M	5.86G	85.17	80.37
SCA-Gate (Ours)	30.60M	5.00G	87.00	86.61

Key Findings

• SCA-Gate is the largest single contributor: Adding SCA-Gate on top of the established BT + CA + RB baseline boosts Synapse Dice from 85.19 to 87.00 (+1.81), vastly exceeding all alternative attention gating options. The joint design of multiplicative consistency + spatial competition is the core engine of performance gains.
• RA is indispensable in CA: Without RA, CA only reaches 79.38 Synapse Dice. Adding RA elevates the performance under the same configuration to 83.57 (+4.19), proving that global residual attention plays a decisive role in the quality of multi-level semantic aggregation; simple averaging is far from sufficient for generating beneficial global context.
• Deep Supervision + Edge Supervision are complementary: Radar charts manifest that the coverage area is maximized (largest Dice outer ring, smallest HD95 inner ring) when both supervisions are simultaneously enabled. ES is particularly helpful for boundary precision; enabling ES alone achieves more significant HD95 improvements than enabling DS alone.
• Stage 0 has low cost-effectiveness: Porting an extra refinement path from the raw input brings only a minor 0.21 boost in Synapse Dice, but spikes GFLOPs from 5.0G to 8.3G (+66%). Consequently, it is excluded from the final model.
• Encoder compatibility is thoroughly validated: Swapping to ConvNeXt-Base (93.72M, 86.79 Synapse) and MaxViT-Base (84.19M, 87.57) runs successfully, indicating that performance is positively correlated with encoder power. Under high-resolution inputs (\(512\times512\)), pvt_v2_b2 can achieve 87.99, but FLOPs skyrocket to 67.5G.

Highlights & Insights

• Empirical verification of "decoder is the bottleneck" is exceptionally strong: Utilizing the identical PVTv2-B2 encoder, the naive decoder yields only 73.91 Synapse Dice, whereas the MedCAGD decoder achieves 87.00. This 13-point performance gap is entirely attributable to the decoder side, which is far more convincing than typical "our refined XX module yields +0.5" style ablations.
• SCA-Gate's multiplicative consistency is a highly transferable design paradigm: Using element-wise multiplication instead of addition to model features of two distinct domains implicitly instills a prior that "only co-activated signals are reliable". This methodology is not only appropriate for skip connections but can also migrate to multi-modal VLMs for vision-language alignment, cross-frame temporal feature fusion in video, etc.—wherever "consistency checking" between two feature maps is needed, multiplication can be tested as an alternative to addition.
• ECA-MSP serves as a plug-and-play near-zero-cost enhancement: Channel attention with multi-scale pooling introduces merely two extra pooling branches (requiring virtually zero additional parameters), yet transitions channel attention from "global-only" to "global-local joint". This improvement can be directly adopted by any architecture utilizing SE/ECA, with minimal migration cost.
• The progressive design of the ablation study is exemplary: The sequential ablation setup of BT \(\rightarrow\) +CA(w/o RA) \(\rightarrow\) +CA(w/ RA) \(\rightarrow\) +RB \(\rightarrow\) +SCA-Gate not only highlights the marginal contribution of each module, but also exposes the significance of sub-components through the "w/ vs w/o RA" contrast in CA—this nested ablation is much more informative than binary "with vs without" comparisons in pinpointing core mechanisms.

Limitations & Future Work

• Only validated on 2D segmentation: All experiments rely on \(224\times224\) 2D inputs, excluding 3D medical images (CT/MRI volume data). In 3D segmentation, decoders face greater challenges in cross-layer alignment and computational efficiency—CA would require interpolation alignment across more levels, and the spatial competition of SCA-Gate would grow cubically in computation.
• OOD robustness is not evaluated: The benchmark does not assess generalization capabilities under domain shifts such as cross-institution, cross-device, or cross-population data. In clinical deployment, this is often a more critical bottleneck than absolute accuracy.
• Incomplete comparison against SAM-like foundation models: In Table 1, AutoSAM and Medical SAM3 have sparse coverage on some datasets, and Medical SAM3 only obtains 55.80 on DRIVE (likely due to insufficient adaptation to fine-grained fundus vessel tasks), diminishing the persuasiveness of the comparison.
• Single-minded choice of BCE Loss may limit boundary performance: The HD95 on Synapse is inferior to MCADS (14.39 vs 11.11), possibly stemming from BCE being less direct at penalizing boundary pixels than Dice Loss or boundary-aware losses (e.g., Boundary Loss). Combining BCE and Dice is standard practice in medical segmentation and remains to be explored.
• Future Directions: (1) Extending SCA-Gate to a multi-head version, enabling different "heads" to focus on diverse spatial competitive patterns (e.g., one head for organ boundaries and another for homogeneous inner structures); (2) validating the transferability of the decoder design inside the 3D framework of nnU-Net; (3) exploring the integration of CA's global context mechanism with prompt-based segmentation (treating global context as implicit prompts).

Related Work & Insights

• vs EMCAD (CVPR 2024): Also a decoder-focused segmentation approach. EMCAD performs multi-stage feature enhancement utilizing refined CBAM-like attention in the decoder. The core distinction of MedCAGD is deploying multiplicative gating (SCA-Gate) instead of additive attention for skip fusion, alongside its unique multi-scale spatial competition design. MedCAGD achieves superior Dice with lower FLOPs (5.0G vs 5.6G).
• vs MCADS: MCADS leverages residual linear attention + depth-to-space upsampling to retain fine-grained structures, exhibiting outstanding HD95 performance (obtaining the best Synapse score of 11.11). However, MCADS requires more parameters (50.90M vs 30.60M), and MedCAGD surpasses it in Dice and IoU, implying better overall segment quality and regional overlap.
• vs UCTransNet: Employs learnable Transformer channel fusion to substitute fixed skip connections, showing a similar motivation to SCA-Gate. SCA-Gate can be perceived as a lighter substitute—utilizing channel attention + depthwise convolutions for selective fusion, yet dodging the quadratic computational complexity of self-attention. Highly apt for resource-constrained scenarios.
• vs SAM-like methods: SAM relies heavily on prompt interactions and mammoth adaptation data, lagging behind customized structures in fixed-task medical segmentation. The results of MedCAGD further empower the "dedicated lightweight models vs general-purpose foundation models" debate in clinical domains—proving that a well-designed specialized decoder is more effective than larger general-purpose models, at least on established benchmarks.

Rating

• Novelty: ⭐⭐⭐⭐ — Although the thesis "the decoder is the bottleneck" is not entirely new, the joint design of multiplicative consistency + multi-scale spatial competition in SCA-Gate offers substantial innovation beyond simple piling of existing attention mechanisms; ECA-MSP represents a small but practical advancement.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — 11 datasets cover 7 imaging modalities, 2D and 3D (ACDC uses 2D slices of 3D MRI), binary and multi-class segmentation. The validation matrix is virtually flawless: progressive component ablations + independent skip attention comparisons + encoder substitutions + resolution analysis + DS/ES supervision signal ablations.
• Writing Quality: ⭐⭐⭐⭐ — Detailed methodology, intuitive equations and illustrations, and each module is backed by matching ablation results; Related Work is slightly verbose but covers comprehensive breadth.
• Value: ⭐⭐⭐⭐ — Achieving SOTA across 11 benchmarks with a lightweight footprint (30.60M / 5.0G) represents high pragmatic value. Both SCA-Gate and ECA-MSP possess powerful reuse potential, easily transplanting as plug-and-play modules into general computer vision tasks.

Related Papers

• [CVPR 2025] Decoding Matters: Efficient Mamba-Based Decoder with Distribution-Aware Deep Supervision for Medical Image Segmentation
• [CVPR 2025] Show and Segment: Universal Medical Image Segmentation via In-Context Learning
• [AAAI 2026] Towards Effective and Efficient Context-aware Nucleus Detection in Histopathology Whole Slide Images
• [CVPR 2026] MedCLIPSeg: Probabilistic Vision-Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation
• [AAAI 2026] Decoding with Structured Awareness: Integrating Directional, Frequency-Spatial, and Structural Attention for Medical Image Segmentation