Implicit neural representation for medical image reconstruction

Medical image reconstruction aims to generate high-quality images from incompletely sampled raw sensor data, which poses an ill-posed inverse problem. Traditional iterative reconstruction methods rely on prior information to empirically construct regularization terms, a process that is not trivial. While deep learning-based supervised reconstruction has made significant progress in improving image quality, it requires large-scale training data, which is difficult to obtain in medical imaging. Recently, implicit neural representation (INR) has emerged as a promising approach, offering a flexible and continuous representation of images by modeling the underlying signal as a function of spatial coordinates. This allows INR to capture fine details and complex structures more effectively than conventional discrete methods. This paper provides a comprehensive review of INR-based medical image reconstruction techniques, highlighting its growing impact on the field. The benefits of INR in both image and measurement domains are presented, and its advantages, limitations, and future research directions are discussed.