Abstract:

In view of the problems such as small size, heterogeneous shape and fuzzy details of the fluid accumulation area, this study integrates CNN(Convolutional Neural Networks) and Transformer to propose an innovative multi-branch segmentation network. The network consists of full convolutional path, Transformer path and CNN-Transformer fusion path. The fully convolutional path is used to capture detailed features of the lesion area, while the Transformer path extracts multi-scale non-local feature information with long-range dependencies. The fusion path takes advantage of both CNN and Transformer to make up for the shortcomings of other branches. The features of the three branches are integrated through the prediction head to generate the final segmentation map. The performance of retinal effusion segmentation is tested on Kermany, UMN and DUKE data sets for intraretinal effusion and subretinal effusion. The experimental results show that the Dice coefficient of the proposed method is 86.63%, the crossover ratio is 77.02%, the sensitivity is 89.47%, and the accuracy is 85.51%, which proves its effectiveness and provides a feasible solution for the automatic segmentation of retinal effusion.

Key words: retinal OCT images, convolutional neural network, Transformer, segmentation network, IRF, SRF, retinal effusion, attention mechanism