Difficulty-Aware Attention Network with Confidence Learning for Medical Image Segmentation
Authors: Dong Nie, Li Wang, Lei Xiang, Sihang Zhou, Ehsan Adeli, Dinggang Shen1085-1092
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on clinical and challenge datasets show that our proposed network can achieve state-of-the-art segmentation accuracy. |
| Researcher Affiliation | Academia | 1Department of Computer Science, University of North Carolina at Chapel Hill, NC 27514, USA 2Department of Radiology and BRIC, University of North Carolina at Chapel Hill, NC 27514, USA 3Institute for Medical Imaging Technology, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China 4College of Computer, National University of Defense Technology, Changsha, China 5Stanford University, Stanford, CA 94305, USA |
| Pseudocode | No | No pseudocode or clearly labeled algorithm blocks were found in the paper. |
| Open Source Code | No | Pytorch1 is adopted to implement our proposed framework shown in Fig. 1. 1https://github.com/pytorch/pytorch - This link refers to the PyTorch framework, not the authors' specific implementation code. No explicit statement of code release for their work was found. |
| Open Datasets | Yes | The first dataset is our own pelvic dataset and the other two are both publicly available challenge datasets which will be introduced in later subsections. ... We further validate our proposed method on MR Brain dataset2. This dataset contains 7 subjects, each with T1 MRI, Flair and manually labeled ground truth map. The task is to segment each voxel into one of the following (tissue) types: background, cortical gray matter (CGM), basal ganglia (BG), white matter (WM), WM lesion (WML), cerebrospinal fluid in the extracerebral space (CSF), ventricle (V), cerebellum (C), brain stem (BS), infarction, and other. ... We also evaluate our proposed method on the prostate segmentation challenge dataset whose ground-truth label maps are hidden from the participants. (The footnotes provide URLs for these public datasets: 2http://mrbrains18.isi.uu.nl/ and 3https://promise12.grand-challenge.org/evaluation/results/.) |
| Dataset Splits | Yes | Five-fold cross-validation is used to evaluate our method. Specifically, in each fold of cross-validation, we randomly chose 35 subjects as the training set, 5 subjects as the validation set, and the remaining 10 subjects as the testing set. |
| Hardware Specification | Yes | A Titan X GPU server is utilized to train the networks. |
| Software Dependencies | No | Pytorch1 is adopted to implement our proposed framework shown in Fig. 1. However, no specific version number for PyTorch or any other software dependencies is provided. |
| Experiment Setup | Yes | The input size of the segmentation network is 64 64 16. The network weights are initialized by the Xavier algorithm (Glorot and Bengio 2010) and weight decay is set to be 1e-4. For the network biases, we initialize them to 0. The learning rates for the segmentation network and the confidence network are both initialized to 5e-3, followed by decreasing the learning rate 2 times for the S, and 5 times for the D every 3 epochs during the training. |