GraphCL: Graph-based Clustering for Semi-Supervised Medical Image Segmentation

Authors: Mengzhu Wang, Houcheng Su, Jiao Li, Chuan Li, Nan Yin, Li Shen, Jingcai Guo

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on three standard benchmarks show that the proposed Graph CL algorithm outperforms state-of-the-art semi-supervised medical image segmentation methods. The source code is available at https://github.com/ dreamkily/Graph CL
Researcher Affiliation Academia 1Hebei University of Technology 2Hong Kong University of Science and Technology 3University of Electronic Science and Technology of China 4National University of Defense Technology 5Sun Yat-Sen University 6The Hong Kong Polytechnic University. Correspondence to: Nan Yin <EMAIL>, Jingcai Guo <EMAIL>.
Pseudocode No The paper describes the methodology using prose and mathematical equations in sections such as '2. Method', '2.1. Notations and Definitions', '2.2. Bidirectional Copy-Paste Framework', and '2.3. Structural Graph Model for Segmentation', but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at https://github.com/ dreamkily/Graph CL
Open Datasets Yes All experiments are performed on three public datasets with different imaging modalities and segmentation tasks: Automatic Cardiac Diagnosis Challenge dataset (ACDC) (Bernard et al., 2018), Atrial Segmentation Challenge dataset (LA) (Xiong et al., 2021) and Pancreas-NIH dataset (Roth et al., 2015).
Dataset Splits Yes Following the protocol in SS-Net (Wu et al., 2022), we conduct semi-supervised experiments with different labeled data ratios (i.e., 5% and 10%). For Pancreas-NIH dataset, we evaluate with a labeled ratio of 20% (Luo et al., 2021a; Shi et al., 2021).
Hardware Specification Yes LA Dataset experiments run on an NVIDIA A800 GPU, while Pancreas-NIH and ACDC datasets use an NVIDIA 3090 GPU.
Software Dependencies No The paper mentions the use of optimizers like SGD and Adam, and network backbones such as 3D V-Net and 2D U-Net, but does not provide specific version numbers for any software libraries or dependencies (e.g., PyTorch, TensorFlow, CUDA, Python).
Experiment Setup Yes All experiments use default settings of α = 0.5, κ = 0.01 and τ = 2, with fixed random seeds. LA Dataset experiments run on an NVIDIA A800 GPU... Training uses SGD with an initial learning rate of 0.01, decaying by 10% every 2.5K iterations. We adopt a 3D V-Net backbone, with patches cropped to 112 × 112 × 80 ... Batch size is 8, split equally between labeled and unlabeled patches, with pre-training and self-training at 5K and 15K iterations, respectively.