Empirical Study of Deep Neural Network Architectures for Non-Rigid Medical Image Registration

Medical image registration is the alignment of two or more images of the same scene or object, but taken possibly from different viewpoints, at different times or by different sensors. Accurate registration plays an important role in the diagnosis and treatment of diseases. Several factors make the task of medical image registration challenging. The surface curvature of the tissues implies that the medical image registration is non-rigid and non-linear. Additionally, the quality of acquired images could be poor because of noise, inherent pathologies, low overlap area and repeated patterns. Recent development in computer vision and medical image processing has seen the introduction of transformer-based networks in accomplishing various tasks and with notable results. This trend has been seen in medical image registration where the performance of convolutional-based networks is being challenged by transformer-based networks. However, it is unclear that whether the improvement cited for transformer-based networks is due mainly to the architecture or other factors such as scale of transformation fields, dataset characteristics and the guidance of different loss functions. In this study, several deep neural network architectures are critically reviewed from the viewpoint of components of architectures, loss functions, scale of transformation fields and datasets respectively. Experiments involving ablation studies over several architectural options were designed and conducted to reveal the performance differences. Theoretical analyses are provided to interpret results.