FRCM-Net: A Fourier-Guided Region Confidence Mixing Network for semi-supervised multi-task carotid plaque segmentation and vulnerability classification

This paper proposes Fourier-Guided Region Confidence Mixing Network (FRCM-Net), a semi-supervised multi-task framework for carotid plaque segmentation and vulnerability classification in dual-view ultrasound. A multi-scale Region Confidence Module aggregates global, plaque, and vessel features derived from segmentation maps and decoder features to improve image-level vulnerability classification. Fourier-guided pseudo-label construction measures the stability of unlabeled predictions under low-frequency amplitude perturbations and combines it with prediction confidence to select reliable pixel-level supervision. During CutMix training, pairs of unlabeled images, pseudo-labels, and reliability masks are synchronously mixed to improve segmentation robustness. The resulting Fourier stability maps also serve as soft gates in the classification module to suppress unreliable regions, connecting segmentation reliability with region-aware classification. Experimental results show that FRCM-Net achieves improved segmentation performance and competitive classification accuracy compared with representative semi-supervised and multi-task methods. Ablation studies further demonstrate the effectiveness of Fourier-guided pseudo-label refinement and region confidence modeling in improving the overall multi-task performance.