Unsupervised Machine Learning Medical Image Registration using Recurrent Deformed Pyramids Networks

For many therapeutic uses, including image-guided therapies, picture registration is crucial in medicine. Though challenging, this field of medical picture registration has lately seen significant algorithmic performance improvements because to the rise of machine learning. Some medical applications are made possible by deep neural networks, such as faster and more accurate picture registration, which is crucial in the fight against malignancies during surgery. Medical image registration procedures sometimes include complex distortion issues. Despite the several suggested sophisticated registration models, deformable registration continues to be a tough task, particularly when dealing with large volumetric deformations, in terms of accuracy and efficiency. Our unsupervised pyramid attention networks (PAN) and network are designed for non-rigid and flexible registration, respectively, with this goal in mind. In contrast to networks that depend on high-weight attentions or transformers, our network is a clean convolutional pyramid, making full use of the pyramid structure's inherent benefits. Specifically, our network guarantees the deformation field's rationality by predicting it from coarse to fine using a step-by-step recursion method that incorporates high-level semantics. Our network is able to achieve deformable registration without the need for separate affine pre-alignment because of the recursive pyramid technique. In terms of Dice scores, average symmetric surface separation, Hausdorff distance, and Jacobian, our network constantly outperforms state-of-the-art, according to experimental data. Our network continues to perform well in adjusting for the huge deformation even when the affine pre-alignment is not included in the data.