One-shot domain adaptation in multiple sclerosis lesion segmentation using convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Convolutional neural networks (CNNs) show promise for automated white matter lesion segmentation in multiple sclerosis (MS).
• Existing CNN methods struggle with adaptability to new imaging domains, leading to decreased accuracy on unseen data.
• Domain shift between training and testing datasets is a significant challenge in medical image analysis.

Purpose of the Study:

• To analyze the effect of intensity domain adaptation on a CNN-based MS lesion segmentation method.
• To evaluate the transferability of a pre-trained CNN model to new MRI scanner protocols and datasets.
• To determine the minimum annotated data and re-trained layers required for comparable accuracy in new domains.

Main Methods:

• Investigated a CNN model trained on public MS datasets.
• Evaluated model performance on data from a clinical center and the ISBI2015 challenge.
• Assessed domain adaptation by varying the number of annotated samples and re-trained layers in the target domain.
• Compared the proposed method's domain adaptation capability against state-of-the-art techniques.

Main Results:

• The proposed CNN model effectively adapted previously acquired knowledge to new image domains, even with limited target dataset samples.
• A one-shot domain adaptation model achieved performance comparable to fully trained CNNs on the ISBI2015 challenge.
• The model demonstrated accuracy similar to human expert raters in lesion segmentation after minimal adaptation.

Conclusions:

• The developed domain adaptation approach enhances the robustness and generalizability of CNN-based MS lesion segmentation.
• This method enables accurate lesion segmentation in diverse clinical settings with reduced annotation requirements.
• The findings suggest significant reductions in time and economic costs associated with manual lesion labeling for MS patients.