Artificial Intelligence Algorithm Can Predict Lymph Node Malignancy from Endobronchial Ultrasound Transbronchial
Yogita S Patel1, Anthony A Gatti2,3, Forough Farrokhyar4
1Division of Thoracic Surgery, Department of Surgery, McMaster University, Hamilton, Ontario, Canada, patelys@mcmaster.ca.
Respiration; International Review of Thoracic Diseases
|September 15, 2024
Summary
An artificial intelligence (AI) algorithm can predict lung cancer nodal metastases from ultrasound images with high accuracy and specificity. This AI tool shows promise for improving lung cancer staging via EBUS-TBNA, though further optimization is needed.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Oncology
- Pulmonology
Background:
- Endobronchial ultrasound transbronchial needle aspiration (EBUS-TBNA) for lung cancer staging is operator-dependent, leading to frequent non-diagnostic lymph node (LN) samples.
- This variability impacts the accuracy of lung cancer staging and subsequent treatment decisions.
Purpose of the Study:
- To develop and evaluate an artificial intelligence (AI) algorithm for predicting nodal metastases using only B-mode ultrasound images of lymph nodes obtained during EBUS-TBNA.
- To assess the AI algorithm's performance in identifying malignant lymph nodes compared to final pathology results.
Main Methods:
- A prospectively recorded dataset of 2,569 B-mode ultrasound images from 773 patients undergoing EBUS-TBNA was utilized.
- An ensemble of three deep neural networks (ResNet152V2, InceptionV3, DenseNet201) was trained using transfer learning on 80% of the images.
- The trained ensemble model was applied to the remaining 20% of images (Test Set) and predictions were compared against pathology results.
Main Results:
- The AI ensemble model achieved an overall accuracy of 80.63% in predicting malignancy.
- The model demonstrated high specificity (96.91%) and positive predictive value (85.90%) for detecting nodal metastases.
- Sensitivity was 43.23%, indicating room for improvement in detecting all malignant cases.
Conclusions:
- An AI algorithm capable of identifying nodal metastases from ultrasound images alone has been developed.
- The AI demonstrates good overall accuracy, specificity, and positive predictive value for lung cancer staging.
- Further research with larger datasets is recommended to enhance the algorithm's performance before clinical implementation.
