Artificial Intelligence Algorithm Can Predict Lymph Node Malignancy from Endobronchial Ultrasound Transbronchial Needle Aspiration Images for Non-Small Cell Lung Cancer.
Journal Articles
Overview
Research
Identity
Additional Document Info
View All
Overview
abstract
INTRODUCTION: Endobronchial ultrasound transbronchial needle aspiration (EBUS-TBNA) for lung cancer staging is operator dependent, resulting in high rates of non-diagnostic lymph node (LN) samples. We hypothesized that an artificial intelligence (AI) algorithm can consistently and reliably predict nodal metastases from the ultrasound images of LNs when compared to pathology. METHODS: In this analysis of prospectively recorded B-mode images of mediastinal LNs during EBUS-TBNA, we used transfer learning to build an end-to-end ensemble of three deep neural networks (ResNet152V2, InceptionV3, and DenseNet201). Model hyperparameters were tuned, and the optimal version(s) of each model was trained using 80% of the images. A learned ensemble (multi-layer perceptron) of the optimal versions was applied to the remaining 20% of the images (Test Set). All predictions were compared to the final pathology from nodal biopsies and/or surgical specimen. RESULTS: A total of 2,569 LN images from 773 patients were used. The Training Set included 2,048 LNs, of which 70.02% were benign and 29.98% were malignant on pathology. The Testing Set included 521 LNs, of which 70.06% were benign and 29.94% were malignant on pathology. The final ensemble model had an overall accuracy of 80.63% (95% confidence interval [CI]: 76.93-83.97%), 43.23% sensitivity (95% CI: 35.30-51.41%), 96.91% specificity (95% CI: 94.54-98.45%), 85.90% positive predictive value (95% CI: 76.81-91.80%), 79.68% negative predictive value (95% CI: 77.34-81.83%), and AUC of 0.701 (95% CI: 0.646-0.755) for malignancy. CONCLUSION: There now exists an AI algorithm which can identify nodal metastases based only on ultrasound images with good overall accuracy, specificity, and positive predictive value. Further optimization with larger sample sizes would be beneficial prior to clinical application.
