Improving skull-stripping for infant MRI via weakly supervised domain adaptation using adversarial learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Skull-stripping is crucial for brain MRI analysis.
• Domain shift between adult and newborn brain MRI hinders deep learning model transferability.
• Previous unsupervised domain adaptation methods required no labeled newborn data.

Purpose of the Study:

• To improve skull-stripping for newborn brain MRI by expanding training data with weakly labeled scans.
• To validate the model's generalization across diverse domains: adult, synthetic, public newborn, and private newborn MRI.
• To enhance the robustness and performance of deep learning models for infant brain imaging analysis.

Main Methods:

• Expanded training and validation datasets using weakly labeled newborn MRI from dHCP, a private dataset, and synthetic GMM data.
• Utilized a previously developed unsupervised domain adaptation framework with a similar core model architecture.
• Validated model generalization across four distinct domains.

Main Results:

• The proposed approach demonstrated improved performance and robustness compared to prior methods.
• Achieved a Dice coefficient of 0.9509±0.0055 and Hausdorff distance of 3.0883±0.1833 for newborn MRI data.
• Outperformed state-of-the-art models like SynthStrip (Dice =0.9412±0.0063, Hausdorff =3.1570±0.1389).

Conclusions:

• Incorporating broader training data under weak supervision significantly impacts newborn brain imaging analysis.
• Weakly labeled newborn data improves model performance and generalization capabilities.
• The developed method is effective for newborn brain imaging analysis, offering superior results to existing techniques.