Topology-Aware Parameter-Efficient Adaptation for Cross-Dataset Retinal Vessel Segmentation

Yongsong Huang; Tomo Miyazaki; Kai Xu; Xiaofeng Liu; Yaohou Fan; Shinichiro Omachi

cs.CVAug 10, 2026

Topology-Aware Parameter-Efficient Adaptation for Cross-Dataset Retinal Vessel Segmentation

Authors: Yongsong Huang, Tomo Miyazaki, Kai Xu, Xiaofeng Liu, Yaohou Fan, Shinichiro Omachi

Organizations: Tohoku University · Yale University

Abstract

Retinal vessel segmentation in multi-domain deployment requires a source model to adapt to domains that differ in imaging conditions and annotation conventions. Conventional parameter-efficient fine-tuning reduces target-specific storage, but its highly restricted adaptation subspace can be insufficient for reconstructing thin, connected vascular structures. We therefore ask how target-specific capacity should be allocated so that topology-aware supervision remains effective under a strict per-domain parameter budget. Based on this principle, we propose TAPDecoderFT, a topology-responsive, role-structured adaptation framework. Specifically, TAPDecoderFT shares a fixed source parameter state across deployment domains, uses low-rank residuals for target-specific private/fusion feature mixing, and retains a trainable dense-reconstruction path comprising the decoder, output head, and refinement module. To promote structurally faithful predictions, the compact target state is jointly optimized with a region-overlap and topology-aware objective that encourages centerline continuity and thin-branch recovery. It improves both DSC and clDice over GenericLoRA-r4 and narrow TAP-r4 in all six directions and is comparable to full fine-tuning.

Figures & tables

Figure 1: Multi-target deployment and shared-state adaptation. (a) FullFT maintains a complete parameter vector $\vartheta_{m}^{\star}$ for each target; the inset contrasts region overlap with thin-vessel connectivity. (b) TAPDecoderFT reuses a fixed shared state $\theta_{\mathrm{sh}}$ and learns a compact state $\phi_{m}$ for each $\mathcal{T}_{m}$ under region-overlap and topology-sensitive supervision.

Figure 2: Layer-level parameterization of TAPDecoderFT for target domain $t$ . Colors distinguish the frozen shared state $\theta_{\mathrm{sh}}=\{\theta_{E},\theta_{A}^{0}\}$ from the trainable target state $\phi_{t}=\{\omega_{t},\psi_{t},\eta_{t}\}$ ; dashed links denote the target-specific normalization state $b_{t}$ . The flow proceeds through low-rank private/fusion adaptation, decoder reconstruction, and prototype refinement. Only $\phi_{t}$ is optimized, while $b_{t}$ is excluded from parameter accounting.

Figure 3: Direction-level structural trade-off relative to FullFT (positive values favor TAPDecoderFT). TAPDecoderFT achieves higher clDice in five transfer directions while maintaining comparable DSC in most settings. DRIVE $\rightarrow$ STARE (D $\rightarrow$ S) exhibits the largest observed appearance shift.

Figure 4: Qualitative transfer examples. Each row shows a local vessel region and masks from the ground truth, direct-trained reference, transfer FullFT, and TAPDecoderFT; the examples highlight how TAPDecoderFT preserves thin-branch continuity and reduces local fragmentation under cross-dataset adaptation.

Explore similar work

CGDD-Net: Context-Guided Dynamic Detail Modeling for Retinal Vessel Segmentation
Oct 4, 2026Xincheng Li, Xiaoqi Sheng, Xinyu ZhangRetinal Vessel Segmentation Multi-Scale Feature Fusion
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
TAPE: A two-stage parameter-efficient adaptation framework for foundation models in OCT-OCTA analysis
Apr 6, 2026Xiaofei Su, Zengshuo Wang, Minghe Sun +2Vision Foundation Model Adaptation Medical Image Segmentation
Nankai University Optometry & Vision Science Institute, National Key Laboratory of Intelligent Tracking and Forecasting for Infectious Diseases, Tianjin Key Laboratory of Intelligent Robotic, Institute of Robotics and Automatic Information System, Nankai University, Tianjin, 300350, China · Institute of Intelligence Technology and Robotic Systems, Shenzhen Research Institute of Nankai University, Shenzhen, 518083, China
U-PEN Mamba: Progressive Expansion with Selective State-Space Modeling for Efficient Retinal Vessel Segmentation
Sep 21, 2026Abel A. Reyes-Angulo, Sidike Paheding, Vijayan K. Asari +2Mamba Retinal Vessel Segmentation
Michigan Technological University, Houghton, MI, USA · Fairfield University, Fairfield, CT, USA · University of Dayton, Dayton, OH, USA +1

Oct 4, 2026J/K move · Enter open · S save

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

Xincheng Li, Xiaoqi Sheng, Xinyu Zhang

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

Retinal Vessel Segmentation Multi-Scale Feature Fusion Convolutional Neural Networks Deformable Convolution

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

TAPE: A two-stage parameter-efficient adaptation framework for foundation models in OCT-OCTA analysis

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