EDLDNet: Efficient Dual-Line Decoder Segmentation

• The paper proposes a dual-line decoding strategy that leverages a noisy branch as a training-time regularizer to enhance segmentation robustness.
• It employs a Pyramid Vision Transformer encoder and multi-scale convolutional attention modules to achieve high accuracy with reduced computational complexity.
• Empirical results on Synapse, ACDC, SegThor, and LCTSC demonstrate significant improvements in Dice score and mIoU while dramatically reducing MACs.

Searching arXiv for the EDLDNet paper and closely related segmentation papers mentioned in the source data. Efficient Dual-Line Decoder Network (EDLDNet) is a medical image segmentation architecture introduced for multi-organ segmentation under a specific design objective: improving segmentation accuracy without incurring the high computational cost often associated with recent state-of-the-art models. It is presented for organs-at-risk segmentation in settings such as radiation therapy, surgical planning, and diagnostic decision-making, where both precision and efficiency are operationally important. The method combines a Pyramid Vision Transformer v2 encoder, a dual-line decoder, Multi-Scale convolutional Attention Modules, Attention Gates, Up-Convolution Blocks, and a mutation-based loss. Its defining characteristic is that the second decoder branch is used as a noisy, training-time regularizer, while inference uses only the noise-free branch, thereby preserving low test-time computation (Hassan et al., 23 Aug 2025).

1. Problem formulation and design objective

EDLDNet is proposed against a recurring trade-off in medical image segmentation: high-performing architectures often require high computational complexity, whereas lightweight models frequently lose segmentation fidelity. The paper situates this problem in multi-organ and organs-at-risk segmentation, where delineation quality has downstream implications for radiotherapy planning and other clinical workflows (Hassan et al., 23 Aug 2025).

The architecture is explicitly designed to address this tension rather than optimizing only one side of it. The stated motivation is that existing CNN-based and Transformer-based systems, including U-Net variants and recent cascaded or attention-based approaches, often either demand high Multiply-Accumulate Operations (MACs) or sacrifice segmentation quality. EDLDNet is therefore framed as a model that shifts some complexity into training while constraining inference-time cost.

A common misconception in this design space is that multi-branch decoders necessarily imply higher deployment cost. EDLDNet is structured specifically to avoid that outcome: the full dual-line system is active during training, but only the noise-free decoder is executed at inference. This suggests that the second branch is intended less as an ensemble component than as a regularization mechanism embedded into the optimization process.

2. Global architecture and dual-line decoding

The network uses a PVTv2-b2 encoder to extract hierarchical features from the input image. The encoder produces four feature maps, denoted $E_1, E_2, E_3, E_4$, which are then routed into two parallel decoder branches. These branches are architecturally identical except for the treatment of the deepest encoder feature (Hassan et al., 23 Aug 2025).

The first branch is the noise-free decoder. For this branch, the deepest decoded representation is initialized as:

$$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$

The second branch is the noisy decoder, which receives structured perturbation:

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$

and

$$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$

Here, $j=1$ denotes the noise-free decoder and $j=2$ the noisy decoder. The perturbation is applied at the encoder-decoder interface to the deepest encoder feature. The paper’s interpretation is that this noisy branch encourages the model to learn more robust and noise-tolerant representations during training.

The recursive decoding stages are defined as:

$$atn_i^j = AG(E_i, UCB(D_{i+1}^j)), \quad i=3,2,1;\; j=1,2$$

$$D_i^j = MSCAM(UCB(D_{i+1}^j) \oplus atn_i^j), \quad i=3,2,1;\; j=1,2$$

At each stage, the previous decoded feature is upsampled, the corresponding skip feature is filtered through an attention gate, the two are fused by element-wise addition, and the result is refined through MSCAM. This yields a decoder topology in which skip transmission, spatial recovery, and attention-based feature refinement are tightly coupled.

The final segmentation predictions are taken from the last two stages of each decoder:

$$I_i^j = \begin{cases} UpScale_4(Conv_{1\times1}(D_i^j)), & i=1 \ UpScale_8(Conv_{1\times1}(D_i^j)), & i=2 \end{cases}$$

producing the four masks

$I = \{I_1^1, I_2^1, I_1^2, I_2^2\}.$

Although four masks are generated during training, inference uses only $$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$0, the final output of the noise-free branch. The architectural consequence is that robustness-oriented supervision is added without permanently increasing the deployed decoder path.

3. Feature refinement modules

Three decoder-side modules constitute the main feature-processing machinery: MSCAMs, UCBs, and AGs. These components are repeated across decoder stages and are presented as the main architectural enhancements beyond the dual-line structure (Hassan et al., 23 Aug 2025).

MSCAM (Multi-Scale Convolutional Attention Module) is defined as:

$$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$1

It is therefore a sequential composition of a Channel Attention Block (CAB), a Spatial Attention Block (SAB), and a Multi-Scale Convolution Block (MSCB).

The CAB emphasizes informative channels through max-pooling and average-pooling branches followed by pointwise convolutions:

$$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$2

$$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$3

The SAB emphasizes spatially relevant locations using channel-wise max and average pooling, concatenation, and a $$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$4 convolution:

$$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$5

The MSCB introduces multi-scale depth-wise convolutions at $$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$6, $$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$7, and $$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$8, followed by channel shuffle:

$$D_4^{j} = \begin{cases} MSCAM(E_4), & j = 1 \end{cases}$$9

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$0

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$1

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$2

Taken together, CAB, SAB, and MSCB are intended to improve feature representation while remaining efficient. The paper attributes to MSCAM the ability to capture local detail and broader contextual information, which is particularly relevant when organ shapes and boundaries vary substantially.

UCB (Up-Convolution Block) performs learned upsampling:

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$3

Its function is to restore spatial resolution while maintaining feature quality. In the ablation study on Synapse, the UCB-only configuration achieves 81.08% Dice, indicating that the upsampling path itself contributes materially to segmentation quality.

AG (Attention Gate) refines skip connections by suppressing irrelevant encoder activations and emphasizing useful regions. The paper defines:

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$4

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$5

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$6

The operational role of AG is to filter skip information before fusion with decoder features. In ablation, adding AG increases Dice from 81.08% to 81.92% on Synapse. This does not make AG the dominant contributor in isolation, but it indicates a measurable gain from skip-selection mechanisms.

4. Mutation-based supervision and training protocol

A central aspect of EDLDNet is its mutation-based loss, which aggregates supervision across combinations of outputs from both decoder branches (Hassan et al., 23 Aug 2025). Rather than supervising only a single prediction, the method forms all non-empty subsets of the four masks:

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$7

so that the powerset used in training is

$$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$8

Because there are four predictions, all non-empty subsets produce $$\mathcal{N}(x) = x + \epsilon, \quad \epsilon \sim U(-0.3, 0.3)$$9 unique mutated predictions. For each mutated prediction, the loss is:

$$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$0

with $$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$1 and $$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$2. The total objective is:

$$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$3

The paper presents this as a combinatorial loss aggregation strategy. Its stated purpose is to encourage consistency across coarse-to-fine outputs from both decoders and thereby improve robustness, generalization, and resistance to overfitting. A plausible implication is that the loss couples branch diversity with output agreement: the noisy branch introduces perturbation-driven variation, while the mutation-based objective constrains that variation to remain label-consistent.

The reported training configuration is:

Item	Setting
Optimizer	AdamW
Learning rate	$$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$4
Weight decay	$$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$5
Batch size	6
Epochs	300 for Synapse and ACDC; 200 for SegThor and LCTSC
Input size	$$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$6 for Synapse and ACDC; $$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$7 for SegThor and LCTSC
Software	Python 3.8; PyTorch 1.11.0+cu113
Hardware	NVIDIA V100 GPU

The reported evaluation metrics are Dice, mIoU, ASD, and MACs. The inclusion of MACs is important because the paper’s central claim concerns the accuracy-efficiency trade-off, not segmentation quality alone.

5. Datasets and empirical performance

The empirical study spans four public medical imaging datasets covering abdominal CT, cardiac MRI, and thoracic CT (Hassan et al., 23 Aug 2025).

Dataset	Composition
Synapse	abdominal CT, 30 cases, 8 organs, 18 train / 12 test
ACDC	cardiac MRI, 70 train / 10 val / 20 test, 3 organs
SegThor	thoracic CT, 32 train / 8 test, 4 organs
LCTSC	lung CT, 36 train / 12 test / 12 val, 5 organs

On Synapse, EDLDNet reports 84.00% Dice and 75.03% mIoU. The paper contrasts this with UNet’s 54.77G MACs and 70.11% Dice, whereas EDLDNet uses 5.6G MACs. The reduction relative to UNet is given as 89.7% in MACs, while the Dice improvement is 13.89 percentage points. The comparison to PVT-EMCAD-B2 is also explicit: both are reported at 5.6G MACs, but EDLDNet reaches 84.00% Dice and 75.03% mIoU, compared with 83.63% Dice and 74.65% mIoU.

Organ-wise on Synapse, EDLDNet is reported as best on aorta (89.12%), gallbladder (74.15%), liver (95.56%), and stomach (83.99%), while remaining highly competitive on harder organs such as the pancreas and spleen.

On ACDC, EDLDNet achieves 92.25% Dice, ahead of PVT-EMCAD-B2 at 92.12%. The model scores 91.20% on the right ventricle and 89.75% on the myocardium, while being slightly below EMCAD on the left ventricle (95.82% vs. 96.02%). The reported result indicates a strong overall balance across structures rather than uniform dominance on every class.

On SegThor, EDLDNet reports Dice values of 0.82 for esophagus, 0.96 for heart, 0.92 for trachea, and 0.95 for aorta. It also reports ASD values including 0.74 for esophagus and 0.42 for trachea. The paper characterizes the method as either best or tied-best across these organs.

On LCTSC, the reported Dice values are 0.74 for esophagus, 0.89 for spine, 0.94 for heart, 0.97 for left lung, and 0.97 for right lung. The paper also notes strong IoU and ASD values. This supports the claim that the model generalizes beyond abdominal CT to cardiac and thoracic imaging.

6. Ablation, interpretation, and significance

The ablation study on Synapse isolates the contribution of the decoder modules and the noisy branch (Hassan et al., 23 Aug 2025). Starting from the UCB-only configuration, the model achieves 81.08% Dice. Adding AG increases performance to 81.92%. Adding MSCAM yields 82.86%. Combining UCB + AG + MSCAM produces 83.60%. Enabling the noisy decoder yields the full 84.00%.

These results are important for interpreting the architecture. First, the gains are cumulative rather than redundant: each module contributes an increment. Second, the noisy decoder produces the final improvement from 83.60% to 84.00%, which the paper treats as evidence that the branch is not merely auxiliary. The reported pattern suggests that the noisy branch acts as an implicit regularizer that improves generalization.

A second point of interpretation concerns efficiency. EDLDNet’s main claim is not that dual decoding is computationally free, but that its extra cost is confined to training. The inference path excludes the noisy branch and uses only the noise-free prediction $$D_4^{j} = \begin{cases} MSCAM(\mathcal{N}(E_4)), & j = 2 \end{cases}$$8. This directly addresses the misconception that regularization through architectural duplication must also be paid for at deployment time.

The broader significance attributed to EDLDNet is that strong multi-organ segmentation accuracy can be obtained without increasing inference-time complexity, provided that additional structure is introduced into training through dual-line decoding and multi-output mutation supervision. The architecture combines a transformer encoder, efficient attention-based multi-scale decoding, and noise-regularized training, and the reported results across Synapse, ACDC, SegThor, and LCTSC are presented as evidence of robustness, computational efficiency, and generalization. The paper further states that the source code, pre-processed data, and pre-trained weights will be available at the project repository linked in the manuscript.