CGDD-Net: Context-Guided Dynamic Detail Modeling for Retinal Vessel Segmentation

Xincheng Li; Xiaoqi Sheng; Xinyu Zhang

cs.CVOct 4, 2026

CGDD-Net: Context-Guided Dynamic Detail Modeling for Retinal Vessel Segmentation

Authors: Xincheng Li, Xiaoqi Sheng, Xinyu Zhang

Organizations: School of Computer Science, The University of Auckland, Auckland, 1010, New Zealand · School of Future Technology, South China University of Technology, Guangzhou, 510641, Guangdong, China

Abstract

Accurate retinal vessel segmentation requires features that capture vascular geometry while preserving information for fine-scale reconstruction. We propose CGDD-Net, a context-guided dynamic detail modeling network that connects adaptive feature extraction to a shared decoder pathway. Context-Guided Scale-Adaptive Deformable Encoding (CSDE) combines fixed-grid convolution with deformable local attention to capture vascular patterns at multiple spatial extents. Spatially Adaptive Multi-Kernel Gating (SAMG) selects receptive-field responses at each location. Dynamic Cross-Scale Detail Fusion (DCDF) aligns the gated intermediate features and compresses them into an eight-channel representation, which is reused at three decoder resolutions together with selected encoder skips. This design consolidates intermediate information before decoding instead of transferring each middle-stage feature through a separate direct skip. On DRIVE, CHASE_DB1, STARE, and HRF, CGDD-Net achieves AUC values of 0.9824, 0.9938, 0.9895, and 0.9874, with F1 scores of 0.8323, 0.8102, 0.8510, and 0.8157, respectively. The complete model contains 1.96 million trainable parameters. In cumulative ablations, the full model improves F1 over the internal baseline by 2.48, 0.61, and 3.49 percentage points on DRIVE, CHASE_DB1, and STARE. Twelve directed cross-dataset experiments further characterize transfer without target-domain adaptation. The results support shared intermediate detail delivery as an effective, parameter-compact architecture for retinal vessel segmentation. Code is available at https://github.com/lixincheng-xcl/CGDD-Net.

Figures & tables

Fig. 1: Representative image structures motivating CGDD-Net: (a) tortuous vessels and bifurcations, (b) coexisting thick and thin vessels, and (c) low-contrast thin vessels. From top to bottom, each column shows the fundus image, the region indicated by the red box, and its manual vessel annotation (GT).

Fig. 2: CGDD-Net architecture with a shared intermediate detail pathway. CSDE generates encoder features $E_{1}$ – $E_{5}$ . SAMG selects multi-kernel responses from $E_{2}$ , $E_{3}$ , and $E_{4}$ , which DCDF fuses into $F_{\mathrm{detail}}$ at quarter input resolution with $C_{d}=8$ . Magenta dashed arrows route this feature to D3, D2, and D1 through independent projections; red dashed arrows denote the retained E4 and E1 skips. A single criss-cross attention (CCA) block follows D3 fusion. CBR denotes convolution, batch normalization, and ReLU. The final $1\times 1$ convolution produces vessel logits.

Fig. 3: Context-Guided Scale-Adaptive Deformable Encoding (CSDE). The upper branch extracts the fixed-grid feature $F_{l}$ using two $3\times 3$ CBR blocks. The lower branch produces $F_{a}$ through deformable local attention, large-kernel depthwise convolution, and pointwise refinement. Its four attention heads use base spacings $\{1,3,5,7\}$ with nine offset-adjusted sampling locations per head. The branch outputs are concatenated, projected from $2C$ to $C$ channels, and added to the input. Circled C and $\oplus$ denote channel concatenation and element-wise addition, respectively; CBR denotes convolution, batch normalization, and ReLU.

Fig. 4: Spatially Adaptive Multi-Kernel Gating (SAMG). Parallel CBR branches extract responses with kernel sizes

1

3

5

, and

7

. Pooled global logits are broadcast and added to local logits, then normalized by softmax across the four branches. The resulting weights vary by location and are shared across channels. Weighted summation produces

F_{s}^{d}

without changing the stage resolution or channel width. Here,

\odot

denotes channel-broadcast multiplication; CBR denotes convolution, batch normalization, and ReLU.

Fig. 5: Dynamic Cross-Scale Detail Fusion (DCDF). Gated features $F_{2}^{d}$ , $F_{3}^{d}$ , and $F_{4}^{d}$ are aligned to the $E_{3}$ grid by downsampling, identity mapping, and upsampling, respectively. Independent $1\times 1$ CBR projections map each feature to $C_{d}=8$ channels. Concatenation followed by a $3C_{d}$ -to- $C_{d}$ projection produces the shared feature $F_{\mathrm{detail}}$ at quarter input resolution. CBR denotes convolution, batch normalization, and ReLU.

Dataset	Quantity	Resolution	Train - Test
DRIVE	40	$565\times 584$	20 - 20
CHASE_DB1	28	$999\times 960$	20 - 8
STARE	20	$700\times 605$	5 - Fold
HRF	45	$3504\times 2336$	25 - 20

Table 1: An overview of our datasets.

Metrics	Calculation formula
SE (Sensitivity)	$\mathrm{SE}=\mathrm{TP}/(\mathrm{TP}+\mathrm{FN})$
SP (Specificity)	$\mathrm{SP}=\mathrm{TN}/(\mathrm{TN}+\mathrm{FP})$
ACC (Accuracy)	$\mathrm{ACC}=(\mathrm{TP}+\mathrm{TN})/(\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN})$
F1 (F1 score)	$\mathrm{F1}=(2\times\mathrm{TP})/(2\times\mathrm{TP}+\mathrm{FP}+\mathrm{FN})$
AUC	Area Under the ROC Curve.

Table 2: Metrics for evaluating the performance.

Methods	SE	SP	ACC	F1	AUC
U-Net [ 35 ]	0.7537	0.9820	0.9531	0.8142	0.9755
LadderNet [ 50 ]	0.7856	0.9810	0.9561	0.8202	0.9793
CE-Net [ 12 ]	0.8209	0.9751	0.9547	0.8202	0.9795
DEU-Net [ 38 ]	0.7940	0.9816	0.9567	0.8270	0.9772
Xu [ 42 ]	0.8339	0.9750	0.9571	0.8319	0.9821
MAU-Net [ 22 ]	0.7890	0.9799	0.9557	0.8192	0.9774

Table 3: Comparison on DRIVE. Best and second-best values per metric are highlighted in red and cyan, respectively; literature results follow the protocols of the original studies.

Methods	SE	SP	ACC	F1	AUC
U-Net [ 35 ]	0.8288	0.9701	0.9578	0.7783	0.9772
LadderNet [ 50 ]	0.7978	0.9818	0.9656	0.8031	0.9839
CE-Net [ 12 ]	0.8027	0.9813	0.9651	0.8012	0.9827
DEU-Net [ 38 ]	0.8074	0.9821	0.9661	0.8037	0.9812
Xu [ 42 ]	0.8508	0.9780	0.9651	0.8270	0.9873
Iter-Net [ 24 ]	0.7970	0.9823	0.9655	0.8073	0.9851

Table 4: Comparison on CHASE_DB1. Best and second-best values per metric are highlighted in red and cyan, respectively; literature results follow the protocols of the original studies.

Methods	SE	SP	ACC	F1	AUC
U-Net [ 35 ]	0.6749	0.9886	0.9576	0.7655	0.9737
LadderNet [ 50 ]	0.7010	0.9895	0.9598	0.7817	0.9809
CE-Net [ 12 ]	0.8296	0.9841	0.9667	0.8296	0.9867
HANet [ 39 ]	0.8186	0.9844	0.9673	0.8379	0.9881
Iter-Net [ 24 ]	0.7715	0.9886	0.9701	0.8146	0.9881
CS 2 -Net [ 31 ]	0.8396	0.9813	0.9670	0.8420	0.9875

Table 5: Comparison on STARE. Best and second-best values per metric are highlighted in red and cyan, respectively; literature results follow the protocols of the original studies.

Method	SE	SP	ACC	F1	AUC
U-Net [ 35 ]	0.7794	0.9804	0.9641	0.7967	0.9721
DUNet [ 19 ]	0.7464	0.9874	0.9651	–	0.9831
HANet [ 39 ]	0.7803	0.9843	0.9653	0.8074	0.9837
MFI-Net [ 44 ]	0.7544	0.9884	0.9668	0.8054	0.9858
M2U-Net [ 21 ]	–	–	0.9635	0.7814	–
DPN [ 14 ]	0.7926	0.9764	0.9591	0.7835	0.9697

Table 6: Comparison with existing methods on HRF. Best and second-best values in each metric column are highlighted in red and cyan, respectively; literature results follow the evaluation protocols reported by the original studies. A dash denotes an unreported metric, and the M2U-Net F1 entry is its reported binary Dice score.

Fig. 6: Qualitative vessel-segmentation comparison. Dataset groups are ordered from top to bottom as DRIVE, STARE, CHASE_DB1, and HRF. Columns show the fundus image, U-Net3+, U-Net, LadderNet, ConvUNeXt, AG-Net, CGDD-Net, and ground truth (GT). Each group contains a full-image view followed by matched enlargements of the yellow- and blue-boxed regions. The crops allow direct comparison of narrow branches, vessel boundaries, and separation between adjacent vessels.

Method	DRIVE AUC	$p$ -value	CHASE_DB1 AUC	$p$ -value	STARE AUC	$p$ -value
U-Net [ 35 ]	0.9755 $\pm$ 0.0008	$5.26\times 10^{-9}$ *	0.9772 $\pm$ 0.0010	$1.51\times 10^{-9}$ *	0.9737 $\pm$ 0.0006	$4.76\times 10^{-23}$ *
LadderNet [ 50 ]	0.9793 $\pm$ 0.0014	$1.64\times 10^{-2}$ *	0.9839 $\pm$ 0.0010	$5.45\times 10^{-5}$ *	0.9809 $\pm$ 0.0009	$2.81\times 10^{-9}$ *
CE-Net [ 12 ]	0.9795 $\pm$ 0.0013	$1.98\times 10^{-2}$ *	0.9827 $\pm$ 0.0018	$1.32\times 10^{-6}$ *	0.9867 $\pm$ 0.0010	$4.32\times 10^{-2}$ *
MPS-Net [ 26 ]	0.9805 $\pm$ 0.0014	$5.43\times 10^{-2}$	0.9869 $\pm$ 0.0009	$3.50\times 10^{-2}$ *	0.9873 $\pm$ 0.0012	$1.41\times 10^{-2}$ *
MFI-Net [ 44 ]	0.9836 $\pm$ 0.0004	$7.75\times 10^{-2}$	0.9879 $\pm$ 0.0005	$5.55\times 10^{-1}$	0.9887 $\pm$ 0.0007	$7.90\times 10^{-3}$ *
CS 2 -Net [ 31 ]	0.9784 $\pm$ 0.0007	$2.09\times 10^{-5}$ *	0.9851 $\pm$ 0.0008	$4.79\times 10^{-4}$ *	0.9875 $\pm$ 0.0012	$1.56\times 10^{-4}$ *

Table 7: AUC summaries and reported Wilcoxon rank-sum comparisons on DRIVE, CHASE_DB1, and STARE. AUC is given as mean

\pm

standard deviation; ∗ denotes a reported

p<0.05

for comparison with CGDD-Net. Red and cyan identify the two highest means in each dataset. Section 4.4 states the descriptive use of these summaries.

Model	DRIVE					CHASE_DB1					STARE
	SE	SP	ACC	F1	AUC	SE	SP	ACC	F1	AUC	SE	SP	ACC	F1	AUC
Baseline	0.7537	0.9820	0.9531	0.8075	0.9801	0.8063	0.9820	0.9607	0.8041	0.9817	0.7810	0.9778	0.9630	0.8161	0.9684
+ CSDE	0.7974	0.9822	0.9672	0.8139	0.9805	0.8220	0.9824	0.9655	0.8059	0.9821	0.7965	0.9802	0.9655	0.8271	0.9729
+ SAMG	0.8157	0.9824	0.9681	0.8201	0.9816	0.8527	0.9789	0.9713	0.8063	0.9839	0.8431	0.9731	0.9669	0.8373	0.9817
+ DCDF	0.8227	0.9823	0.9688	0.8258	0.9821	0.8466	0.9859	0.9725	0.8086	0.9870	0.8462	0.9817	0.9695	0.8363	0.9859
+ Detail-Guided Decoder	0.8265	0.9818	0.9697	0.8301	0.9823	0.8532	0.9877	0.9734	0.8197	0.9887	0.8470	0.9860	0.9720	0.8424	0.9867

Table 8: Cumulative ablation study on DRIVE, CHASE_DB1, and STARE. Each row builds on the preceding configuration, following the progression described in Section 4.5 . The final CGDD-Net row additionally includes the CCA refinement after D3 fusion. The best and second-best values in each metric column are highlighted in red and cyan, respectively.

Fig. 7: Qualitative comparison of the cumulative configurations in Table 8 . Columns show (a) the fundus image, (b) the baseline, (c) +CSDE, (d) +SAMG, (e) +DCDF, (f) +Detail-Guided Decoder, (g) +Selective Skip, (h) the full CGDD-Net with CCA, and (i) ground truth. Configurations (c)–(h) cumulatively extend the preceding model. The top row shows the full-image results; the middle and bottom rows magnify the yellow- and blue-boxed regions, respectively.

Dataset	F1	ACC	AUC	SE	SP
D $\rightarrow$ C	71.44 ( -9.58 )	96.11 ( -1.41 )	99.21 ( -0.17 )	79.64 ( -6.36 )	97.11 ( -1.04 )
D $\rightarrow$ S	82.49 ( -2.61 )	97.25 ( -0.50 )	98.90 ( -0.05 )	87.27 ( +0.66 )	97.59 ( -0.53 )
D $\rightarrow$ H	77.61 ( -3.96 )	96.11 ( -1.00 )	98.90 ( +0.16 )	87.84 ( +4.22 )	96.80 ( -1.43 )
C $\rightarrow$ D	68.94 ( -14.29 )	95.67 ( -1.37 )	80.05 ( -18.19 )	55.39 ( -28.82 )	99.26 ( +1.58 )
C $\rightarrow$ S	76.61 ( -8.49 )	97.00 ( -0.75 )	86.25 ( -12.70 )	66.28 ( -20.33 )	99.10 ( +0.98 )
C $\rightarrow$ H	79.06 ( -2.51 )	96.90 ( -0.21 )	96.11 ( -2.63 )	76.24 ( -7.38 )	98.65 ( +0.42 )

Table 9: Direct zero-shot cross-dataset performance of CGDD-Net. Each entry reports the target score (%) followed by its difference from the corresponding target-domain reference (percentage points). No target-domain fine-tuning, threshold tuning, or normalization adaptation is used. D=DRIVE, C=CHASE_DB1, S=STARE, and H=HRF.

Fig. 8: Hyperparameter sensitivity on CHASE_DB1. Accuracy (ACC) is plotted against (a) dropout rate, (b) batch size, and (c) learning rate, varying one factor while keeping the remaining settings fixed. Red markers identify the configuration used in the main experiments: dropout

0.1

, batch size

64

, and learning rate

10^{-3}

. Labels above the markers give the corresponding ACC values.

Method	AUC (%)	F1 (%)	Parameters (M)	FLOPs (G)
U-Net [ 35 ]	97.37	76.55	34.53	50.11
AttU-Net [ 32 ]	97.75	75.43	34.88	50.96
IterNet [ 24 ]	98.81	81.46	1.36	19.44
R2U-Net [ 1 ]	97.61	77.48	39.09	117.08
AMFEF-Net [ 49 ]	98.39	79.85	36.79	52.33
MDF-Net [ 30 ]	99.36	87.03	2.83	46.36

Table 10: Model complexity and segmentation performance on STARE. AUC and F1 are percentages. Parameter counts and FLOPs are reported in millions and billions, respectively. Red and cyan identify the best and second-best values in each column. FLOPs are retained as reported model-forward counts, not as measurements of full-image latency under a common profiling configuration.

Fig. 9: Probability and binary outputs for a low-contrast retinal image. From left to right: the fundus image, the CGDD-Net probability map, its binary segmentation at threshold $0.5$ , the ConvUNeXt binary segmentation, and ground truth. Brighter values in the probability map indicate stronger vessel responses. Paired crops below each image enlarge the red- and yellow-boxed regions, allowing faint responses to be inspected before and after thresholding.

Explore similar work

Sep 21, 2026cs.CV

U-PEN Mamba: Progressive Expansion with Selective State-Space Modeling for Efficient Retinal Vessel Segmentation

Accurate retinal vessel segmentation is important for computer-aided ophthalmic analysis, yet thin vessels, low contrast, and severe foreground-background imbalance remain challenging for encoder-decoder networks. This paper presents U-PEN Mamba, a U-shaped retinal vessel segmentation architecture that couples progressive nonlinear feature expansion with selective state-space modeling. The proposed network enriches local vessel responses with progressive expansion, models long-range spatial dependencies through a Mamba Global Context (MGC) block with linear sequence complexity, and uses attention-based decoder fusion to recover fine vascular boundaries. We evaluate U-PEN Mamba on CHASE DB1 and DRIVE using a consistent patch-based preprocessing pipeline and compare it with convolutional, attention-based, transformer-based, and Mamba-based segmentation baselines. U-PEN Mamba obtains the best mean intersection over union among the compared methods, achieving 0.8394 on CHASE DB1 and 0.8221 on DRIVE, with Dice scores of 0.8187 and 0.8078, respectively, using 21.6M trainable parameters. Ablation studies show that the MGC block contributes the largest gain over the U-Net baseline, while projection dimension and state size provide practical accuracy-efficiency control. These results indicate that selective state-space modeling is a promising global-context mechanism for parameter-efficient retinal vessel segmentation. Code is available at: https://github.com/areyesan/UPEN_Mamba.

Abel A. Reyes-Angulo, Sidike Paheding, Vijayan K. Asari +2

Michigan Technological University, Houghton, MI, USA · Fairfield University, Fairfield, CT, USA · University of Dayton, Dayton, OH, USA +1

Vessel Segmentation U-Net Topomamba Mamba Vessels State Space Models

Ask PDF

Jun 3, 2026cs.CV

LightVesselNet: An Ultra-Lightweight Sub-100K Parameter Network for Retinal Blood Vessel Segmentation

Retinal blood vessel segmentation plays a vital role in the early detection of diabetic retinopathy and glaucoma. While recent deep learning models have achieved great segmentation accuracy, they typically require heavy computational resources, making real-world deployment on edge devices difficult. In this paper, we propose LightVesselNet, an efficient neural network designed for retinal vessel segmentation in a resource-constrained environment. Despite containing only 75K parameters, LightVesselNet performs competitively with much larger models. The network employs a compact encoder decoder architecture enhanced with channel and spatial attention mechanisms, a multi-scale feature aggregation module at the bottleneck, and a subpixel upsampling strategy in the decoder. A dedicated edge residual connection preserves fine vessel detail throughout decoding. Extensive experiments on five publicly available datasets: DRIVE, STARE, CHASEDB1, FIVES, and HRF, yield sensitivity scores of 0.8189, 0.8499, 0.8640, 0.8634, 0.8096, and Dice coefficients of 0.8070, 0.8072, 0.8181, 0.8649, and 0.7686, respectively. LightVesselNet shows improved efficiency (Performance vs Parameter or GFlops) compared to State-of-the-Art models. Cross-dataset evaluation confirms the model's generalisation capability. Overall, LightVesselNet is a strong candidate for deployment in low-resource clinical settings and mobile screening tools.

Shadman Sobhan, Farhana Jalil

Department of Electrical & Electronic Engineering, Bangladesh University of Engineering and Technology (BUET), Dhaka, Bangladesh.

Vessel Segmentation Mobilenetv2 Deep Learning Upsampling

Ask PDF

May 12, 2024eess.IV

Explainable Convolutional Neural Networks for Retinal Fundus Classification and Cutting-Edge Segmentation Models for Retinal Blood Vessels from Fundus Images

Early detection of vision-threatening conditions such as diabetic retinopathy, glaucoma, and age-related macular degeneration depends on retinal fundus image analysis, but manual assessment is slow and expert-dependent. Automated convolutional neural networks classify fundus images accurately yet act as black boxes, and existing retinal vessel segmentation methods lose discriminative power under pathology and seldom exploit attention or transformer backbones. Using the FIVES and DRIVE fundus datasets, we develop a two-pipeline framework that pairs four-class disease classification with attention- and transformer-based vessel segmentation, organised in three stages: (1) FIVES images are augmented by rotation and horizontal and vertical flips and used to fine-tune eight ImageNet-pretrained CNNs: ResNet101, DenseNet169, Xception, InceptionV3, DenseNet121, InceptionResNetV2, ResNet50, and EfficientNetB0. (2) Five gradient-based explanation methods, Grad-CAM, Grad-CAM++, Score-CAM, Faster Score-CAM, and Layer-CAM, are computed on the final convolutional block of each classifier and compared qualitatively across architectures. (3) Ten U-Net variants are benchmarked for vessel segmentation: TransUNet (hybrid CNN--Transformer encoder) and Attention U-Net (gated skip connections), evaluated with ResNet50V2, ResNet101V2, and ResNet152V2 backbones, along with additional Attention U-Net configurations using DenseNet backbones, and the fully transformer-based Swin-UNet. ResNet101 gives the highest classification accuracy: 94.17% (F1 0.942)

>

88.33% for EfficientNetB0. For segmentation, the architecture ranking is consistent on both datasets: Attention U-Net

>

TransUNet

>

Swin-UNet. The strongest configuration is Attention U-Net with a ResNet101V2 backbone: FIVES IoU 0.722, Dice 0.838; DRIVE IoU 0.648, Dice 0.787, lifting DRIVE IoU 60.80

\rightarrow

64.83 over a prior custom U-Net.

Fatema Tuj Johora Faria, Mukaffi Bin Moin, Pronay Debnath +2

aDepartment of Computer Science and Engineering, Ahsanullah University of Science and Technology, Dhaka, Bangladesh

Vessel Segmentation Fundus Images Retinal Imaging Semantic Segmentation Convolutional

Ask PDF