Single-Domain Generalization in Medical Image Segmentation via Test-Time Adaptation from Shape Dictionary
Authors: Quande Liu, Cheng Chen, Qi Dou, Pheng-Ann Heng1756-1764
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two medical image segmentation tasks demonstrate the consistent improvements of our method across various unseen domains, as well as its superiority over state-of-the-art approaches in addressing domain generalization under the worst-case scenario. |
| Researcher Affiliation | Academia | Quande Liu1, Cheng Chen1, Qi Dou1, Pheng-Ann Heng1,2 1 Department of Computer Science and Engineering, The Chinese University of Hong Kong 2 Guangdong-Hong Kong-Macao Joint Laboratory of Human-Machine Intelligence-Synergy Systems, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences |
| Pseudocode | No | The paper describes its methods in text and diagrams but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | No | The paper does not provide any statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | Prostate MRI segmentation: We collect prostate T2-weighted MRIs from six different clinical centers out of three public datasets, including NCI-ISBI13 (Bloch et al. 2015), I2CVB (Lemaˆıtre et al. 2015) and PROMISE12 (Litjens et al. 2014) datasets. Fundus image segmentation: We employ retinal fundus images from four medical institutions out of three public datasets, including REFUGE (Orlando et al. 2020), Drishti GS (Sivaswamy et al. 2015) and RIM-ONE-r3 (Fumero et al. 2011) datasets. |
| Dataset Splits | No | The paper describes using a single source domain for training and unseen target domains for testing, but it does not specify explicit train/validation/test splits within these datasets or typical percentages for data partitioning as commonly done with a validation set. |
| Hardware Specification | Yes | The framework is implemented with Pytorch using one NVIDIA Titan Xp GPU. |
| Software Dependencies | No | The paper mentions 'Pytorch' but does not specify its version number or any other software dependencies with their versions. |
| Experiment Setup | Yes | In our implementation, the element number K in the shape dictionary is set as 48. For input perturbations, two dropout layers with dropout rate 0.5 are added before the bottleneck layer and the last convolution layer respectively, and Gaussian noise with magnitude 0.1 is added onto the input images. The step size for test-time update is 1e-3. We employ an adapted Mix-residual-UNet (Liu et al. 2020) as segmentation backbone. The standard data augmentation techniques are used to avoid overfitting, including random rotation and flipping (horizontal and vertical). The model is trained using Adam optimizer with momentum of 0.9 and 0.99, and the learning rate is initialized as 1e-3. We totally train 100 epochs on the single source domain as the network has converged, with batch size set as 5. |