Neural networks for registration based on perceptual loss

Deep learning has emerged as a powerful tool for accurate and efficient registration of medical images. However, traditional evaluation methods based on Euclidean distance (MSE) or normalized cross-correlation (NCC) suffer from low contrast or large deformation of the image. To address these issues, inspired by the concept of Large Language Model (LLM), this work trains a feed-forward neural network on multi-site medical image datasets to construct the Perceptual Loss which is applied to the registration process replacing the traditional loss functions. The method of ours is validated on the LPBA40 and improves the Dice value by 2% compared to the second ranked method. At the same time, faster convergence and smoother deformation fields are achieved by this method, which adapts to handling complex deformations effectively. Our code can be found at https://github.com/nicetry0724/smilecode.git.