Transfer learning by feature-space transformation: A method for Hippocampus segmentation across scanners.

Area of Science:

• Medical Imaging
• Neuroimaging
• Computer Vision

Background:

• Supervised voxel classification is a common MR brain segmentation method.
• Performance degrades when training and test data have different acquisition parameters (e.g., scanners).
• Differences in feature representations cause performance deterioration.

Purpose of the Study:

• To propose a feature-space transformation (FST) method.
• To overcome performance degradation in MR brain segmentation due to scanner variability.
• To improve the robustness of segmentation models.

Main Methods:

• Developed a feature-space transformation (FST) using unlabeled images from dual-protocol scans.
• Utilized affine registration to establish a feature-space mapping between source and target voxels.
• Mapped training samples to the feature space of test samples.

Main Results:

• The proposed FST significantly improved hippocampus segmentation performance on datasets with both small and large differences between training and test images.
• FST outperformed methods relying solely on image normalization.
• FST enhanced the performance of a state-of-the-art patch-based-atlas-fusion technique.

Conclusions:

• Feature-space transformation is an effective method to address scanner variability in MR brain segmentation.
• The proposed FST improves segmentation accuracy and robustness.
• This approach offers a valuable solution for multi-center or longitudinal studies using MR imaging.