A multiple organ segmentation system for CT image series using Attention-LSTM fused U-Net

• Published in: Multimedia tools and applications Vol. 81; no. 9; pp. 11881 - 11895
• Main Authors: Chen, Pin-Hsiu, Huang, Cheng-Hsien, Chiu, Wen-Tse, Liao, Chen-Mao, Lin, Yu-Ruei, Hung, Shih-Kai, Chen, Liang-Cheng, Hsieh, Hui-Ling, Chiou, Wen-Yen, Lee, Moon-Sing, Lin, Hon-Yi, Liu, Wei-Min
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2022; Springer Nature B.V
• Subjects: 1195: Deep Learning for Multimedia Signal Processing and Applications; Accuracy; Computed tomography; Computer Communication Networks; Computer Science; Contouring; Data processing; Data Structures and Information Theory; Dosage; Esophagus; Gallbladder; Graphical user interface; Image acquisition; Image segmentation; Kidneys; Liver; Medical imaging; Multimedia; Multimedia Information Systems; Neural networks; Organs; Physicians; Radiation therapy; Semantics; Special Purpose and Application-Based Systems; Spinal cord; Stomach; Workflow; LSTM; CT image; DICOM-RT; Attention U-net; Contour; Multiple organ segmentation system
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-021-11889-7

The multi-organ contouring (MOC) task is required when a radiotherapy is performed to eradicate the cancerous tissue while minimizing the dosage delivered to the surrounding healthy organs. Currently most of the task is done manually with enormous labor and time cost. To reduce the scheduling waiting time from increasing cancer population, it is beneficial to both the patients and therapeutists to have an automatic contouring tool. In this work an Attention-LSTM fused U-Net model is proposed to perform the multiple organ segmentation from a CT image series. The organs to be delineated include lung, liver, stomach, esophagus, heart, and kidneys. To train and evaluate our model, the CT image series of 146 patients was acquired from a local hospital with IRB approval. The segmentation accuracy of the six organs in terms of Dice Similarity Coefficient (DSC) were 99.27%, 95.48%, 88.53%, 80.81%, 93.8%, and 93.46%, respectively. To make the AI-embedded MOC system readily applicable in clinical environments, a data processing workflow and the corresponding GUI were also implemented and published on Github. The doctors can download the CT image data from the PACS server, use our system to perform MOC tasks, and output the contouring results in DICOM-RT format so they can be uploaded back to the treatment planning system for further fine-tuning and dosage/path calculation. To our best knowledge the work might be the first non-commercial model-integrated system compatible with the commercial treatment planning systems and ready to be used by the doctors.