Inpainting of ultrasound cardiac tissues with consistent anatomical structures

Abstract

Image inpainting techniques play a crucial role in restoring occluded regions in medical images, ultimately enhancing the consistency of clinical diagnoses. Traditional inpainting methods often face challenges when attempting to restore complex anatomical structures and large occluded areas in medical ultrasound images. While recent deep learning-based techniques show promise, they still encounter issues such as loss of edge information and inconsistent reconstruction in occluded regions. In this paper, we introduce a Segmentation-guided Medical Ultrasound Inpainting framework designed to overcome these limitations. Our framework integrates segmentation-derived edge priors to guide the inpainting process, ensuring anatomical consistency and enhancing the recovery of fine details. We propose a Multi-Scale Mixed Residual Block to improve the model’s ability to restore large masked regions and a Deformable Edge Attention mechanism to preserve critical edge details during downsampling while minimizing the introduction of noise. Extensive experiments on two publicly available echocardiography datasets show that our method significantly outperforms state-of-the-art inpainting models in terms of PSNR, MAE, and SSIM metrics. The results highlight the potential of our approach to provide accurate, consistent ultrasound image reconstruction, making it a valuable tool for clinical diagnostics.