AUTHOR=Zheng Minting , Guo Chenhua , Zhu Yifeng , Gang Xiaoming , Fu Chongyang , Wang Shaowu 

TITLE=Segmentation model of soft tissue sarcoma based on self-supervised learning

JOURNAL=Frontiers in Oncology

VOLUME=Volume 14 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2024.1247396

DOI=10.3389/fonc.2024.1247396

ISSN=2234-943X

ABSTRACT=Soft tissue sarcoma refers to sarcomas derived from soft tissues such as smooth muscle, fat, and fibrous tissue its incidence is similar to that of cervical cancer and esophageal cancer. To better segment the sarcoma area in soft tissue imaging and assist clinicians in diagnosing the patient's condition, we collect and process multi-modal MRI images (a total of 8,640 images) of 45 patients with soft tissue sarcoma of the thigh (COR) and combine with several clinicians to label the sarcoma area and obtain a data set of soft tissue sarcoma of the thigh. To verify the validity and availability of the data set, we develop a novel multi-modal segmentation model based on the basic UNet medical segmentation model, using residual networks and attention mechanisms to extract more important information for each modality, and using self-supervised learning strategies to optimize the ability of encoders to extract features. At the same time, considering the different emphasis of different modalities on the characterization of tumor regions, we use the developed model to demonstrate the validity and usability of the data set we created through further experiments. The experimental results show that the model uses multi-modal MRI images as input is better rather than single-modal MRI images as input in the segmentation effect.