Feasibility of artificial intelligence computer aided diagnosis technology for predicting benign or malignant nodules using radial probe endobronchial ultrasonography images.

K. Shikano*^a (Dr), T. Nakaguchi^b (Prof), T. Nakajima^c (Dr), N. Akira^a (Dr), A. Mitsuhiro^a (Dr), K. Takeshi^a (Dr), I. Jun^a (Dr), S. Takuji^a (Prof)

^a Department of Respirology, Graduate School of Medicine, Chiba University, Chiba, JAPAN ; ^b Center for Frontier Medical Engineering, Chiba, JAPAN ; ^c Department of General Thoracic Surgery, Dokkyo Medical University, Tochigi, JAPAN

* kshikano@chiba-u.jp

Background: Patients with peripheral pulmonary nodule requires to be diagnosed by pathology and radial probe endobronchial ultrasonography (RP-EBUS) enabled to confirm the reach to the target and improves the diagnostic yield of transbronchial diagnosis. The image analysis of RP-EBUS provides supplementary information for adequate sampling to the operators. This study aimed to assess the feasibility of artificial intelligence computer aided diagnosis (AI-CAD) technology for predicting benign or malignant nodules using RP-EBUS images.

Methods: We retrospectively reviewed 72 bronchoscopy cases using RP-EBUS for diagnosing peripheral pulmonary nodules. The final diagnosis was made by the pathological results and/or clinical follow-up of more than six months. The ultrasound images for training and validation were extracted from the video clips in the MOV format. We used ResNet-18 and ResNet-101 as a deep learning model, and the learning rate was fixed at 0.001. We performed a grid search to explore the optimal condition for AI-CAD by changing the number of training images and batch size.

Results: From 11 benign and 61 malignant nodules, 10,464 images were extracted, including 2,217 benign and 8,247 malignant images, and were analyzed in this study. For the training set, 9 benign and 50 malignant cases were selected, and the benign and malignant images were assigned one to one. Two benign and 11 malignant cases were selected for validation, and every 100 images were tested. When we used ResNet-101, learning data of 500 images and batch size of 64, the highest yield was achieved with a maximum sensitivity of 0.83.

Conclusion: AI-CAD technology could predict the benign or malignant nodule with a maximum sensitivity of 0.83. However, further improvements would be needed; this technology might provide helpful information to bronchoscopists during a transbronchial biopsy for peripheral pulmonary nodule using RP-EBUS.

Disclosure of funding source(s): none