GraphMorph: Tubular Structure Extraction by Morphing Predicted Graphs
Authors: Zhao Zhang, Ziwei Zhao, Dong Wang, Liwei Wang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The efficacy of our method in the centerline extraction and segmentation tasks has been substantiated through experimental evaluations across various datasets. We conduct the experiments by beginning with the centerline extraction task to verify the effects of the two components of Graph Morph. Experimentally, serveing as an auxiliary training module to learn the graph representation, the Graph Decoder enhances the segmentation network s focus on branch-level features, thus both volumetric metrics and topological metrics are boosted. |
| Researcher Affiliation | Collaboration | Zhao Zhang1,5 Ziwei Zhao2 Dong Wang2 Liwei Wang3,4,B 1Center for Data Science, Peking University 2Yizhun Medical AI Co., Ltd 3State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University 4Center for Machine Learning Research, Peking University 5Pazhou Laboratory (Huangpu), Guangzhou, Guangdong, China |
| Pseudocode | Yes | The pseudo-code of our Skeleton Dijkstra algorithm is given in Algorithm 2, which finds the optimal path satisfying the skeleton nature for two points. Algorithm 1 Morph Module |
| Open Source Code | No | Source code will be released soon. |
| Open Datasets | Yes | We evaluate Graph Morph on three medical datasets and one road dataset. DRIVE [39] and STARE [13] are retinal vessel datasets commonly used in medical image segmentation. ISBI12 [1] contains 30 Electron Microscopy images to segment membranes. The Massachusetts Roads (Mass Road) dataset contains 1171 aerial images for road network extraction. |
| Dataset Splits | Yes | For ISBI12, following previous works [35, 29], we split it into 15 images for training and 15 for testing. For Mass Road, we follow [40] to construct the training set, and the total 63 images of the official validation set and test set are used for testing. These cases were divided in a 7:1:2 ratio for training, validation, and testing. |
| Hardware Specification | Yes | Experiments were conducted using an NVIDIA GeForce RTX 3090 with 24 GB GPU memory. The CPU used was an Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, which features 28 cores. |
| Software Dependencies | No | We implement Graph Morph based on Py Torch [31] and Detectron2 [42]. However, specific version numbers for these software dependencies are not provided. |
| Experiment Setup | Yes | For three medical datasets, we use randomly cropped 384 × 384 images for training. The size of ROI samples H is 32 and the stride of sliding window used in inference process is 30. For all experiments, we use 64 ROI samples per image (R = 64) to train the Graph Decoder, and the number of node queries in the modified Deformable DETR is set to 100 (K = 100). According to previous experiences [38], the default hyperparameters used in loss functions are as follows: λclass = 0.2, λcoord = 0.5, α = 0.6, γ = 2. We use ADAM optimizer with the initial learning rate 1e-3 and cosine learning rate schedule with warm-up strategy to train the network. The weight decay is set to be 1e-4 uniformly. We train the network for 3K iterations for the three medical image datasets, and 10K for Mass Road. We use batchsize=4 for all datasets. We use pthresh = 0.5 across all experiments. |