Endobronchial Ultrasound Multimodal Videos Based Artificial Intelligence EBUSNet for Segmentation and Prediction of Intrathoracic Lymphadenopathy

J. Chen^a (Dr), J. Li^b (Mr), C. Gu^a (Ms), X. Zheng^a (Dr), Y. Li^a (Dr), F. Xie^a (Dr), S. Liu^a (Mr), W. Dai^b (Prof), J. Sun*^a (Prof)

^a Department of Respiratory Endoscopy, Department of Respiratory and Critical Care Medicine, Shanghai Chest Hospital, Shanghai Jiao Tong University, Shanghai, CHINA ; ^b School of Electronic Information & Electrical Engineering, Shanghai Jiao Tong University, Shanghai, CHINA

* xkyyjysun@163.com

Junxiang Chen, Jin Li and Chuanjia Gu contributed equally to this article.

Background: Endobronchial ultrasound (EBUS) imaging has been established as a cornerstone in initial detection and evaluation the nature of intrathoracic lymphadenopathy. Nevertheless, the inherent limitation of clinical practice due to the dynamic image related interpretation such as subjectivity of bronchoscopist and interobserver variability. In this study we developed a EBUS network (EBUSnet) for segmentation and prediction the property of intrathoracic lymphadenopathy based on EBUS multimodal videos.

Methods: In the study, 1006 lymph nodes (LNs) of EBUS videos were gathered from a single center between July 2018 and June 2020 retrospectively to train and validate the EBUSnet, which consists of region of interest detection module and real-time auxiliary diagnosis system. Between July and October 2020, 267 LNs of EBUS videos were prospectively gathered from multiple centers for test dataset. A transfer dataset containing 245 lung lesions of EBUS videos from a single center was collected retrospectively.

Results: In the validation and test cohort, the area under the curve (AUC) and accuracy of multimodal EBUSnet were 88.97% (95% CI: 84.13%-93.82%) and 81.20% (95% CI: 74.47%-87.94%), 84.90% (95% CI: 80.00%-89.80%) and 81.27% (95% CI: 76.07%-85.77%) respectively. The capability of multimodal EBUSnet in distinguishing benign and malignant LNs was superior to unimodal, also excellent than the qualitative analysis of experienced interventional pulmonologist. Besides, the multimodal EBUSnet had the tallest AUC of 80.87% (95% CI: 71.05%-90.68%) and accuracy of 87.07% (95% CI: 80.55%-92.04%) in the transfer cohort.

Conclusion: This study demonstrated that EBUSnet has a valuable performance and generalized ability in predicting the malignancy of intrathoracic lymphadenopathy and lung lesions. The EBUSnet may be beneficial to the automated real-time evaluation of targeted LNs, which assisted in EBUS procedure can valuable reducing unnecessary biopsies and shortening examination time.

Disclosure of funding source(s): none