Topology-Aware Uncertainty for Image Segmentation
Authors: Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, Chao Chen
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On various 2D and 3D datasets, our method produces better structure-wise uncertainty maps compared to existing works. Code available at https://github.com/Saumya-Gupta-26/struct-uncertainty |
| Researcher Affiliation | Collaboration | Saumya Gupta1 Yikai Zhang2 Xiaoling Hu1,3 Prateek Prasanna1 Chao Chen1 1Stony Brook University, NY, USA 2Morgan Stanley, NY, USA 3Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, MA, USA |
| Pseudocode | Yes | A Probabilistic DMT In Algo. 1, we provide a pseudocode of our Probabilistic DMT module proposed in Sec. 3.1. Algorithm 1 Probabilistic DMT pseudocode |
| Open Source Code | Yes | Code available at https://github.com/Saumya-Gupta-26/struct-uncertainty |
| Open Datasets | Yes | We evaluate our method on four datasets: DRIVE [65], ROSE [40], ROADS [44] and PARSE 2022 Grand Challenge [39, 70]. |
| Dataset Splits | Yes | The DRIVE dataset contains 2D retinal vasculature;... We use the dataset s predetermined split of 20 training images and 20 test images. For training, we keep aside four randomly-chosen samples as validation, and train on the remaining 16 samples. ROSE... It has a predetermined split of 30 train and 9 test samples... For training, we keep aside four randomly-chosen samples as validation, and train on the remaining 26 samples. |
| Hardware Specification | Yes | We use the Py Torch framework, a single NVIDIA Tesla V100-SXM2 GPU (32G Memory), and a Dual Intel Xeon Silver 4216 CPU@2.1Ghz (16 cores) for all the experiments. |
| Software Dependencies | No | The paper mentions 'Py Torch framework' but does not specify its version number or other key software dependencies with their respective version numbers. |
| Experiment Setup | Yes | The training hyperparameters for our method as well as the baselines are as tabulated in Tab. 4. Our main hyperparameters are u, α, β, γ... We achieve the best ECE when γ = 0.2, α = 2.0, β = 0.01... For all the experiments, we set u = 0.3. |