PatchGAN-guided semi-supervised U-Net for carotid plaque segmentation in ultrasound under limited annotation conditions.
Authors
Affiliations (10)
Affiliations (10)
- Department of Computer Engineering, RAIT, D. Y. Patil Deemed University, Mumbai, Maharashtra, 400 706, India.
- Department of Computer Engineering, RAIT, D. Y. Patil Deemed University, Mumbai, Maharashtra, 400 706, India. [email protected].
- School of Artificial Intelligence, Sai University, Chennai, Tamil Nadu, 603110, India. [email protected].
- Vascular Screening and Diagnostic Centre, University of Nicosia, Nicosia, 1700, Cyprus.
- Department of Radiology, Azienda Ospedaliero Universitaria (A.O.U.), 10015, Cagliari, Italy.
- School of Artificial Intelligence, Sai University, Chennai, Tamil Nadu, 603110, India.
- Stroke Diagnostic and Monitoring Division, AtheroPoint™, Roseville, CA, 95661, USA.
- Department of Innovation, Global Biomedical Technologies, Inc., Roseville, CA, 95661, USA.
- Department of ECE, Idaho State University, Pocatello, ID, 83209, USA.
- Department of Computer Science & Engineering, Symbiosis Institute of Technology, SymbiosisInternational (Deemed University), Nagpur Campus, Nagpur, Maharashtra, 440008, India.
Abstract
Artificial intelligence models are increasingly used for stroke risk evaluation and clinical decision-making. However, the scarcity of expert masks, low contrast, and noise in ultrasound images affect segmentation performance. Our study innovatively integrated an adversarial PatchGAN discriminator into a batch-normalized semi-supervised U-Net generator to enhance carotid plaque segmentation in ultrasound images with limited expert annotation. The discriminator provides patch-level localized feedback to enhance boundary delineation and structural consistency of the predicted masks. The model was trained using a hybrid loss function and self-training strategy. The performance of our framework was compared with a fine-tuned semi-supervised U-Net model. Both frameworks were trained under identical experimental settings using 30% of the labeled data; the remaining unlabeled images were utilized to generate pseudo-labels for data augmentation. A confidence threshold of 0.7 was applied to filter unreliable predictions. We conducted experiments on a dataset of 970 internal carotid artery ultrasound images from the Imperial College of London, UK. The experimental results showed that the proposed framework achieved a Dice coefficient of 86.12 and a Jaccard index of 75.66, outperforming the baseline U-Net (84.22 and 73.12, respectively). The integration of patch-level adversarial feedback into a semi-supervised framework enhanced the segmentation accuracy and improved the reliability of pseudo-labels under limited supervision. The results were comparable to those of state-of-the-art deep learning models, which require a large pool of labeled data.