Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Towards Generalizable Retina Vessel Segmentation with Deformable Graph Priors

Authors: Ke Liu, Shangde Gao, Yichao Fu, Shangqi Gao

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three public benchmarks (CHASE, DRIVE, HRF) show that Graph Seg consistently outperforms existing methods under domain shifts. These results highlight the importance of jointly modeling anatomical topology and image structure for robust generalizable vessel segmentation.
Researcher Affiliation Academia 1 Zhejiang University 2 University of Cambridge EMAIL EMAIL
Pseudocode No The paper describes the methodology using text and mathematical equations, and provides network architecture diagrams, but does not contain explicit pseudocode or algorithm blocks.
Open Source Code Yes Code can be found at github.com/AI4MOL/Graph Seg.
Open Datasets Yes We employ four widely recognized datasets: CHASE [31], HRF [32], DRIVE [33], and STARE [34].
Dataset Splits Yes To evaluate cross-domain generalization, the models are trained on one dataset and tested on the others. We use the training set of one dataset and the test set of the other datasets to evaluate the generalization performance of our method. We follow the training, validation and test data splitting in [11].
Hardware Specification Yes We implement our method using Py Torch and train it on a cluster with NVIDIA A800 GPU. We trained Graph Seg for 2.8 hours on GTX 4090.
Software Dependencies No We implement our method using Py Torch and train it on a cluster with NVIDIA A800 GPU. The batch size is set to 4, and the learning rate is set to 0.001. We use the Adam optimizer with a weight decay of 1e-5.
Experiment Setup Yes The batch size is set to 4, and the learning rate is set to 0.001. We use the Adam optimizer with a weight decay of 1e-5. The model is trained for 500 epochs, and we use early stopping based on the validation loss. The input images are resized to 256 256 pixels, and data augmentation techniques such as random rotation, and flipping are applied during training. For the image decomposition, we adopted the two Res Nets with ten layers, and for the segmentation network, we used the efficient U-Net [35]. For the matching process, we use two two-layer graph convolutional networks (GCN) [47] with 64 channels and a three-layer convolutional network (CNN).