Automated Removal of Caliper Annotations from Thyroid Ultrasound Images

Background: Ultrasound imaging is widely used owing to its safety, accessibility, and real-time capabilities. However, thyroid ultrasound images frequently contain calipers and measurement lines that obscure anatomy and degrade visual fidelity. Although these overlays are clinically useful during acquisition, their presence in stored images can hinder quantitative analysis and bias the performance of downstream artificial intelligence (AI) models. Compared with artifact reduction in CT and MRI, automated caliper removal in ultrasound remains underexplored. Methods: We designed an attention-enhanced Residual U-Net for automated caliper removal, formulated as an image-to-image translation task using paired thyroid ultrasound images acquired before and after caliper placement. The network reconstructs caliper-free images while preserving diagnostically relevant anatomy by combining residual learning with attention-gated skip connections that selectively suppress caliper-related features. A composite loss based on mean squared error and the structural similarity index (SSIM) balances intensity fidelity and structural preservation. The proposed method was compared with multiple convolutional and transformer-based architectures using standard image-quality metrics and qualitative expert review. Results: The proposed model consistently outperformed competing architectures, achieving higher SSIM, PSNR, Pearson correlation coefficient, and edge-based F1 scores, together with reduced MSE, MAE, and RMSE. Visual inspection confirmed effective removal of caliper annotations while preserving lesion boundaries, tissue textures, and anatomical continuity, with no observable loss of diagnostically relevant information. Conclusions: The proposed approach provides a reliable and computationally efficient solution for automated caliper removal in thyroid ultrasound images. By improving dataset quality and image realism, it enhances image usability for clinical review and strengthens the robustness of AI-driven ultrasound analysis.