Unsupervised Domain Adaptation for Anatomical Structure Detection in Ultrasound Images
Authors: Bin Pu, Xingguo Lv, Jiewen Yang, He Guannan, Xingbo Dong, Yiqun Lin, Li Shengli, Tan Ying, Liu Fei, Ming Chen, Zhe Jin, Kenli Li, Xiaomeng Li
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the proposed To Mo-UDA for ultrasound fetal anatomical structure detection, we introduce FUSH2, a new Fetal Ultra Sound benchmark, comprises Heart and Head images collected from Two health centers, with 16 annotated regions. Our experiments show that utilizing topological and morphological anatomy information in To Mo-UDA greatly improves organ structure detection. This expands the potential for structure detection tasks in medical image analysis. |
| Researcher Affiliation | Academia | 1The Hong Kong University of Science and Technology 2Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, Anhui University 3Sichuan Provincial Maternity and Child Health Care Hospital 4Shenzhen Maternity and Child Healthcare Hospital 5Harbin Red Cross Central Hospital 6Hunan University. |
| Pseudocode | Yes | Algorithms 1 and 2 outline the basic procedures of our proposed Topology Knowledge Transfer (TKT) and Morphology Knowledge Transfer (MKT), respectively. |
| Open Source Code | Yes | Datasets and source code are available at https://github.com/xmedlab/To Mo-UDA. |
| Open Datasets | Yes | In summarize, our contributions include: 1. A comprehensive real-world fetal ultrasound dataset from two health centers with 1,978 heart and 1,391 head views, namely FUSH2, is released. ... Datasets and source code are available at https://github.com/xmedlab/To Mo-UDA. ... MMWHS (Zhuang et al., 2019) consists of 20 unpaired MRI and 20 CT volumes... Cardiac UDA was originally proposed in (Yang et al., 2023)... |
| Dataset Splits | Yes | For each dataset, we split it into training, validation, and test sets with a ratio of 7:1:2, respectively. |
| Hardware Specification | No | The paper mentions "various ultrasound devices such as Samsung and Sono Scape" as data collection sources, but does not specify any hardware (GPU, CPU models, etc.) used for running the experiments or training the models. |
| Software Dependencies | No | The paper mentions using "Res Net101" as a feature extractor and "FCOS head" for detection, along with "Stochastic Gradient Descent (SGD) optimizer". However, it does not provide specific version numbers for any software libraries, frameworks (like PyTorch, TensorFlow), or programming languages. |
| Experiment Setup | Yes | We use Res Net101 (He et al., 2016) as our feature extractor. For the detection head, we choose one-stage (Tian et al., 2019) (two-stage (Ren et al., 2015) is shown in the Table A1) detection strategies. During training, we apply the Stochastic Gradient Descent (SGD) optimizer with an initial learning rate of 0.001, a batch size of 6, a momentum of 0.9, and a total of 100 training epochs. We apply proportional scaling, random flipping, and random erasing as preprocessing operations. ... Our experiment found that when both α and β are set to 0.1, it can achieve the best domain adaptation performance. ... In the GNN module, the embedding size is set to d to stay in line with the input (V, E). |