Neural networks for image processing: deep image prior applied to fetal ultrasounds

Abstract

The point of intersection of artificial intelligence and medical imaging opens the door to innovative tools that can improve patient diagnosis, especially in settings with limited data or computational resources. This study explores the use of the Deep Image Prior (DIP), a training-free deep learning method, applied to fetal ultrasound images to improve their quality through denoising and super-resolution techniques. The project has two principal objectives. The first one is to adapt and implement the DIP method in MATLAB, a program that is less commonly used for this technique, to evaluate its feasibility in a different Python environment. The adaptation of one programming language to another one is done in order to make state-of-the-art image processing tools more accessible to engineers in academic and industrial contexts, where MATLAB is widely used. The second objective is to study this specific method applied to fetal ultrasound images, where the quality and the clarity of the image are vital for accurate diagnosis. They are often degraded by noise and low resolution. The project presents the combination of the DIP’s capacity to restore images with no training data with the ultrasound images clarity issue. Through single-image optimization and without needing large datasets, the study demonstrates how DIP can be a suitable solution for image restoration in scenarios where there are limited resources. The results of the project highlight the potential of training-free methods in medical imaging, showing competitive restoration quality and avoiding the limitations of data-intensive models. In addition, the model developed in the project offers an area for future research on other medical imaging problems, where AI can be a great tool for improving their visualization and detection.