Advanced Image Processing Techniques for Ultrasound Images using Multiscale Self Attention CNN

The aim of this research is to enhance the quality of prenatal ultrasound images by addressing common artifacts such as missing or damaged areas, speckle noise, and other types of distortions that can impede accurate diagnosis. The proposed approach involves a novel preprocessing pipeline for prenatal 5th-month ultrasound scan images, which includes three main steps. First, Multiscale Self Attention convolutional neural network (CNN) is used for image inpainting and augmentation to fill missing or damaged areas and generate augmented images for training DL models. Second, Anisotropic Diffusion Filtering is used for speckle noise reduction, and the filter parameters are adapted to local noise characteristics using memory-based speckle statistics. Third, the CNN is trained to estimate local statistics of the speckle noise and adapt filtering parameters accordingly to capture local and global image features. The effectiveness of the proposed approach is evaluated on a prenatal 5th-month ultrasound scan dataset. The results demonstrate that the proposed preprocessing steps significantly improve the quality of ultrasound images and lead to better performance of DL models. The proposed preprocessing pipeline using Multiscale Self Attention CNN for image inpainting and augmentation, followed by Anisotropic Diffusion Filtering and memory-based speckle statistics for speckle noise reduction, can significantly enhance the quality of prenatal ultrasound images and enhance the accuracy of diagnostic models. The approach has potential for broader use in medical imaging applications.