Unsupervised domain adaptation method for segmenting cross-sectional CCA images.

Area of Science:

• Medical imaging
• Deep learning
• Image segmentation

Background:

• Ultrasound image quality issues (attenuation, noise, shadowing) challenge automatic vessel segmentation.
• Deep convolutional neural networks show promise but traditionally require large labeled datasets.
• Acquiring and annotating in-vivo ultrasound data is difficult and time-consuming.

Purpose of the Study:

• To develop an unsupervised domain adaptation method for robust ultrasound vessel segmentation.
• To overcome the need for extensive manual annotation of in-vivo ultrasound data.
• To improve the generalization of deep learning models across different ultrasound data sources.

Main Methods:

• A two-stage model-based, unsupervised domain adaptation approach was employed.
• The network was initially trained on simulated ultrasound images with accurate ground truth.
• Subsequent unsupervised training on unlabeled in-vivo data utilized prior knowledge of elliptical segmentation masks.

Main Results:

• The proposed method significantly improved segmentation performance, achieving a median Dice similarity coefficient of 0.951.
• Performance surpassed that of a domain adversarial neural network (0.922) and a Star-Kalman algorithm (0.942).
• The unsupervised domain adaptation effectively addressed the challenge of limited labeled in-vivo data.

Conclusions:

• Training on simulated data followed by unsupervised domain adaptation on unlabeled in-vivo data is feasible and effective.
• This approach circumvents the need for large manually labeled in-vivo datasets in deep learning for ultrasound.
• The model-based domain adaptation, using shape priors, offers potential for various applications and domains.