ParaTransCNN: Parallelized TransCNN Encoder for Medical Image Segmentation

The convolutional neural network-based methods have become more and morepopular for medical image segmentation due to their outstanding performance.However, they struggle with capturing long-range dependencies, which areessential for accurately modeling global contextual correlations. Thanks to theability to model long-range dependencies by expanding the receptive field, thetransformer-based methods have gained prominence. Inspired by this, we proposean advanced 2D feature extraction method by combining the convolutional neuralnetwork and Transformer architectures. More specifically, we introduce aparallelized encoder structure, where one branch uses ResNet to extract localinformation from images, while the other branch uses Transformer to extractglobal information. Furthermore, we integrate pyramid structures into theTransformer to extract global information at varying resolutions, especially inintensive prediction tasks. To efficiently utilize the different information inthe parallelized encoder at the decoder stage, we use a channel attentionmodule to merge the features of the encoder and propagate them through skipconnections and bottlenecks. Intensive numerical experiments are performed onboth aortic vessel tree, cardiac, and multi-organ datasets. By comparing withstate-of-the-art medical image segmentation methods, our method is shown withbetter segmentation accuracy, especially on small organs. The code is publiclyavailable on https://github.com/HongkunSun/ParaTransCNN.