PW-CM: A Medical Image Segmentation Based on Consistency Model by Using Patches and Wavelet Transforms

As a new trend, generative consistency models are becoming increasingly popular, and some works have used consistency generative models for image segmentation tasks, achieving good results. However, consistency generative models are large in scale and slow in training, occupying a lot of computational resources when training large image datasets. A straightforward idea is to crop the images into patches before inputting them into the model, but this approach loses the global information of an image. This paper aims to use the low-frequency features obtained from wavelet transformations to preserve global information and produce an image patch of the same size as the others. These image patches are then encoded and inputted into the consistency model for training, which significantly reduces the parameter scale of the consistency model. Additionally, experiments have validated that the PW-CM model can also achieve good results in medical image segmentation.