An advanced W-shaped network with adaptive multi-scale supervision for osteosarcoma segmentation. (February 2023)

• Record Type: Journal Article
• Authors: Shuai, Limei; Zou, Wei; Hu, Nan; Gao, Xin; Wang, JiaJun
• Is Part Of: Biomedical signal processing and control. Volume 80(2023)Part 1
• Publication Date: 2023-02
• Subjects: Segmentation -- Deep learning -- Skip connection -- Attentional mechanism -- Deep supervision; 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.2022.104243
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Osteosarcoma is one of the most common malignant bone tumors in adolescents, hence a precise and reliable automatic segmentation method is urgently needed in clinical practice. In this paper, an advanced W-shaped network is proposed for automatic and accurate segmentation of osteosarcoma in computed tomographic images. This deep model is developed based on two cascaded baseline U-Nets where feature maps of the same scales in encoding and decoding paths of both networks are fused in terms of advanced skip connections. Different from simple skip connections in the traditional U-Net which fuse low-level and high-level feature maps directly, the advanced skip connection module learns fine details from low level feature maps before concatenating to the corresponding high-level feature maps. Multiple side outputs are used to supervise the training process of the network. Multi-scale channel attention module is introduced to enable the network learn to suppress the irrelevant side outputs while highlight the useful ones to osteosarcoma tasks. The performance of our method is evaluated on a home-built dataset containing 2303 computed tomographic images of osteosarcoma whose results show that our method outperforms the U-Net and Multiple Supervised Residual Network with improvements of 7.47% and 2.59% in dice similarity coefficient, respectively. Our method also performs better than our previously developed W-Net++ with an improvement of 1.04% in dice similarity coefficient. Highlights: A multi-scale attention block to assign different weights to different side outputs. An advanced skip connection to extract fine details from the low-level feature map. A multi-scale fusion strategy for features from different U-Nets in the cascade.