Transition Net: 2D backbone to segment 3D brain tumor. (May 2022)

• Record Type: Journal Article
• Authors: Liu, Jiahao; Zheng, Jinhua; Jiao, Ge
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: CNN -- Transformer -- 3D Brain Tumor Segmentation -- Transfer learning -- Weighted Region Loss; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103622
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A segmentation method using 2D backbone to segment 3D multimodal brain tumor images is proposed, which can be extended to other 3D tasks. To solve the problem of large differences between the pre-training domain and the target domain, we propose Transition Head and Transition Decoder to transform the intermediate features. Weighted region loss is proposed. This loss improves on Dice loss and Focal loss and can drive the model to actively learn difficult samples. Abstract: Pre-trained models are widely used in 2D deep-learning processing. Training efficient and generalizable 3D pre-training models requires large amounts of expensive labeled 3D images. This leads many current 3D brain tumor image segmentation methods to spend a lot of effort to design encoders with complex and hard-to-understand structures, and designing a lot of experiments to prove the reliability of its structure, making the task of segmentation tedious. To address these problems, this work uses an existing 2D Backbone to segment 3D brain tumor images (Transition Net) to make the segmentation process simple. We use Swin Transformer as the encoder, combined with a decoder constructed by 3D convolution to segment 3D brain tumor images. We designed the Transition Head and Transition Decoder components to solve the cross-domain variation problem. The Transition Head is used to convert the input data into feature maps suitable for Swin Transformer. Transition Decoder is used to transform the multi-scale feature maps extracted by Backbone and fuse them with the features sampled on the CNN in multiple steps to obtain the final segmentation results. To leverage the correlation between different sub-regions and drive the model to actively learn difficult samples, we propose weighted region loss to train the model. Experiments on the BraTS 2019 dataset show that our method achieves Dice scores of 0.7485, 0.8446, and 0.9125 for ET, TC, and WT regions, respectively. The average segmentation performance of our model for the three subregions is more stable than some other methods.