SwinE-UNet3+: swin transformer encoder network for medical image segmentation

• Published in: Progress in artificial intelligence Vol. 12; no. 1; pp. 99 - 105
• Main Authors: Zou, Ping, Wu, Jian-Sheng
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.03.2023; Springer Nature B.V
• Subjects: Artificial Intelligence; Artificial neural networks; Coders; Computational Intelligence; Computer Imaging; Computer Science; Control; Data Mining and Knowledge Discovery; Datasets; Feature extraction; Image segmentation; Mechatronics; Medical imaging; Natural Language Processing (NLP); Pattern Recognition and Graphics; Robotics; Sampling; Short Communication; Swine; Transformers; Vision; UNet3; Tumor segmentation; Convolutional neural network; Patch merging; Swin Transformer
• ISSN: 2192-6352; 2192-6360
• DOI: 10.1007/s13748-023-00300-1

A SwinE-UNet3+ model is proposed to improve the problem that convolutional neural networks cannot capture long-range feature dependencies due to the limitation of receptive field and is insensitive to contour details in tumor segmentation tasks. Each encoder layer of SwinE-UNet3+ uses two consecutive Swin Transformer blocks to extract features, especially long-range features in images. Patch Merging is used for down-sampling between encoder layers. The decoder uses Conv2DTranspose to perform progressive up-sampling and uses convolution operation to aggregate the decoder information after up-sampling and the encoder information through skip connection. The proposed model evaluates the TipDM Cup rectal cancer dataset and the melanoma dermoscopic image ISIC-2017 dataset. Experimental results show that SwinE-UNet3+ model outperforms UNet, UNet++ and UNet3+ models in Dice coefficient, IOU value and Precision evaluation metric.