Digital-Rapid On-site Examination in Endobronchial Ultrasound-Guided Transbronchial Needle Aspiration (DEBUT): a proof of concept study for the application of artificial intelligence in the bronchoscopy suite

Abstract

Endobronchial ultrasound guided transbronchial needle aspiration (EBUS-TBNA) has become the standard of care for sampling mediastinal and hilar lesions and is finding increased acceptance for diagnostic as well as staging purposes [1]. EBUS-TBNA is an expensive procedure due to the high cost of equipment [2]. A repeat procedure in case of an inconclusive outcome adds to the burgeoning expenditure. To circumvent this, rapid on-site examination (ROSE) has been adopted to reduce the number of needle punctures and decrease the requirement for additional procedures [3]. However, ROSE requires presence of a pathologist or cyto-technicians in the bronchoscopy suite.Artificial intelligence (AI), and deep neural networks in particular, have come a long way since their inception. AI is being used for microscopic examination of pathology images, with PAPNET being the most well-known application being used for screening of cervical cancer [4]. To the best of our knowledge, no study has been published to date on the application of AI during ROSE of EBUS-TBNA.We evaluated the performance of an AI model, consisting of an open-sourced convolutional neural network (CNN) using transfer learning, for its ability to accurately classify images of ROSE of EBUS-TBNA smears in the bronchoscopy suite. This study was approved by the institutional ethics committee (IEC number AIIMS/IEC/2020/3214).A total of 441 cytology images of ROSE of EBUS-TBNA smears were collected retrospectively from medical records of patients who underwent bronchoscopy from June 2019 to June 2020. The smears were stained with modified Giemsa stain and images were obtained using Olympus microscope at 400× magnification. Slides were examined by a pathologist who assigned them to classes based on the criteria adopted from ALSHARIF et al. [5], according to which a smear was labelled as adequate if either there were >40 lymphocytes per high power field (HPF) or presence of granulomas or malignant cells were noted; otherwise, it was categorised as inadequate. Consequent to this classification, the smears were categorised into the following four classes: 1) "granulomas" ( presence of one or more granulomas), 2) "adequate lymphocytes" (>40 lymphocytes per HPF), 3) "malignant cells" ( presence of malignant cells), 4) "inadequate" (meets none of the criteria above) (figure 1).All the images were downsized to 224×224 pixel size in JPG format, which is compatible with different open-sourced CNNs. 15% of the images from each category of smears were separated to form a test set (n=66) and rest of the images constituted the training set (n=375) for creation of a trained AI model and its subsequent evaluation. Different open-sourced convolutional neural networks of varying depths were assessed for their performance on the training dataset. The 19-layered VGG19 model released by the Visual Geometry Group (VGG), Oxford University was found to be the best in terms of accuracy and was used subsequently in this study [6]. VGG19 was implemented using Python through PyTorch. We used stoc...

A combination of flexible and rigid bronchoscopy in the successful removal of a residual fish bone from a peripheral bronchus: A case report

“…For instance, Li et al built a bronchoscopy quality-control system based on AI and showed that the supplemental application of the AI system could reduce the differences in the endoscopic skills of doctors with different levels of experience ( 26 ). However, at present, only limited studies have been taken in the application of AI to bronchoscopy ( Supplementary Table S1 ) ( 26 – 31 ), most of which are mainly focused on recognizing tumors ( 27 , 28 , 31 ). Particularly, there is a research gap in the application of AI in complicated FB aspiration.…”

Development and validation of the artificial intelligence (AI)-based diagnostic model for bronchial lumen identification

“…In terms of endoscopy, Wu et al combined a deep CNN with deep reinforcement learning to train a real-time quality improvement system (WINSENSE) for esophageal gastroduodenoscopy. Monitoring blind spots during the examination has become a powerful auxiliary tool to reduce the different skills of endoscopists and improve the quality of daily endoscopies ( 32 ); however, in relation to bronchoscopy, most studies have focused on the identification of the pathological specimens of transbronchial biopsies ( 33 , 34 ). To our knowledge, this is the first study to apply CNN training to identify bronchial anatomy.…”

Artificial intelligence in interventional pulmonology

“…Notably, this automated cropping method enabled the system to deliver diagnoses within a second. Additionally, Asfahan et al [50 ▪ ] evaluated a CNN-based system designed for categorizing cytology smear images obtained during ROSE in EBUS-TBNA. Their classification included adequate samples (>40 lymphocytes per high-power field), inadequate samples, granulomas, and malignant cells.…”