Semi-supervised U-Net with PatchGAN improves carotid plaque segmentation on ultrasound with limited labels

A PatchGAN-discriminator boosted semi-supervised U-Net to a Dice of 86.12 for carotid plaque ultrasound segmentation with only 30% labeled data, vs. 84.22 for standard semi-supervised U-Net (970 images).

• Trained on 970 internal carotid artery ultrasound images with only 30% expert-labeled data, generating pseudo-labels from the rest.
• The proposed model also achieved higher Jaccard index (75.66 vs 73.12).
• Authors report the segmentation accuracy is comparable to state-of-the-art models requiring full annotation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.