Automatic identification and segmentation of mediastinal lymph nodes and blood vessels in endobronchial ultrasound (EBUS) using a deep neural network

O. Ervik*^a (Dr), IE. Tveten^b (Ms), EF. Hofstad^b (Mr), H. Sorger^a (Dr)

^a Nord-Trøndelag Hospital Trust, Levanger Hospital, Clinic of Medicine, and Norwegian University of Science and Technology, Faculty of Medicine, Department of Circulation and Medical Imaging, Levanger, NORWAY ; ^b Department of Health Research, SINTEF Digital, Trondheim, Norway, Trondheim, NORWAY

* oyvind.ervik@ntnu.no

Background

Endobronchial ultrasound transbronchial needle aspiration (EBUS-TBNA) from thoracic lymph nodes is essential in lung cancer investigation, to select patients for curative therapy and avoid wrong treatment decisions. Each lymph node must be repeatedly localized and sampled, challenging the bronchoscopist´s technique. To improve EBUS-TBNA precision, we propose a deep neural network (DNN) that identifies lymph nodes and blood vessels examined with EBUS.

Methods

Nineteen patients referred to Levanger Hospital for EBUS-TBNA were prospectively enrolled. Ultrasound (US) video was recorded from the EBUS processor (Olympus EVIS EXERA III) to a laptop computer. Mediastinal and hilar lymph nodes were imaged by EBUS (Olympus BF UC19OF) and lymph node stations were labeled in real-time by the bronchoscopist according to their anatomical level. Postprocedurally, a bronchoscopy expert selected static images from the EBUS videos and annotated lymph nodes and vessels using a spline segmentation technique. US images with annotations from fifteen patients were used to train a DNN (U-Net). The performance of the network was evaluated on the remaining four patients.

Results

In total, 510 US frames from nineteen patients were annotated. For segmentation of lymph nodes and blood vessels, respectively, the Dice similarity coefficient was 0.73 and 0.69, precision was 0.74 and 0.73, recall was 0.66 and 0.64, and F1 score was 0.70 and 0.68. Average run-time of processing and segmentation per US frame was 93 ± 6 ms on a laptop with CPU and built-in GPU (Intel® UHD Graphics) using OpenVINO for inference.

Conclusion

Segmentation of EBUS lymph nodes and vessels using a DNN was feasible and fast, making real-time automatic labeling possible. We will further aim to improve segmentation quality and develop software for intraoperative labeling of lymph nodes and vessels during EBUS-TBNA. A dataset from a second hospital is currently prepared for method testing and validation.

Disclosure of funding source(s):

This work was funded by the The Liaison Committee for education, research and innovation in Central Norway and Norwegian National Research Center for Minimally Invasive and Image-Guided Diagnostics and Therapy, St. Olavs hospital, in collaboration with NTNU and SINTEF.