Core Concepts
By combining domain generalization (DG) pre-training and test-time adaptation (TTA), the proposed DG-TTA method enables high-quality segmentation of medical images in unseen target domains without access to the source data.
Abstract
The paper presents a method called DG-TTA that combines domain generalization (DG) pre-training and test-time adaptation (TTA) to enable robust out-of-domain medical image segmentation.
DG Pre-training:
The authors propose using the MIND descriptor and GIN augmentation techniques during pre-training on source data to improve model generalization.
Experiments show that DG pre-training can significantly reduce the performance gap when applying pre-trained models to out-of-domain target data, compared to non-DG pre-trained models.
TTA:
During test-time, the authors apply TTA by optimizing the model weights to produce consistent predictions under different spatial augmentations of the target image.
TTA is shown to further improve performance, especially for non-DG pre-trained models, recovering up to 63.9% Dice score on the abdominal segmentation task.
Experiments:
The authors evaluate their DG-TTA approach on various out-of-domain segmentation scenarios, including abdominal, cardiac, and lumbar spine tasks.
Leveraging the large-scale TS dataset, DG-TTA enables high-quality segmentation on unseen datasets without accessing the source data.
The method is shown to be highly modular and can be easily integrated into the popular nnUNet framework.
Overall, the DG-TTA approach provides a powerful tool to obtain accurate medical image segmentation on unseen data, addressing the common challenge of domain shift in real-world clinical applications.
Stats
The dataset dimensions range from 512 x 512 x 85 vox to 512 x 512 x 198 vox, with fields of view from 280 x 280 x 280 mm3 to 500 x 500 x 650 mm3.
The cardiac MRI data has resolutions from 0.80 x 0.80 x 1.00 mm3/vox down to 1.00 x 1.00 x 1.60 mm3/vox.
The lumbar spine MRI data has a resolution of 1.8 x 1.8 x 4.0 mm3/vox.
Quotes
"Applying pre-trained medical segmentation models on out-of-domain images often yields predictions of insufficient quality."
"We argue that linking both approaches [DG and TTA] enables optimal separate use of source and target data where DG maximizes the base performance and TTA can further optimize the result."
"Our method enables separate use of source and target data and thus removes current data availability barriers."