A Learning-Free Approach to Mitigate Abnormal Deformations in Medical Image Registration

Medical image registration is a critical process in several diagnostic and therapeutic procedures. While deep-learning deformable registration models have demonstrated reasonable accuracy, they can produce abnormal deformations that introduce substantial artifacts in medical images by unrealistically modifying the shape and position of anatomical structures. These abnormal deformations may not be effectively detected or mitigated during inference due to their similarity to the natural elastic deformation of soft tissue. Moreover, the limited generalizability of learning-based approaches restricts their ability to assess image registration beyond their training scope. In this paper, we propose a learning-free method for estimating and correcting abnormal deformations, which are responsible for the maximum deformation error. Our proposed model-agnostic approach introduces variations to both the input images of the registration model and the model weights at inference, making it adaptable to a wide range of deep-learning-based medical image registration models. Next, the proposed approach uses the variabilities in the estimated registration deformation fields to mitigate significant deformation errors. We evaluate our proposed approach on two datasets: a 3D abdominal computed tomography dataset (the LiTS dataset), and a 3D brain magnetic resonance imaging dataset (the OASIS dataset) using synthetically generated deformation fields that resembles patient and organ movement as well as changes in organ sizes; reducing the maximum registration error by up to 6.1% for the first and 5.7% for the second. These findings demonstrate that our approach can significantly mitigate abnormal deformations in medical image registration, improving accuracy and reducing artifacts.