Machine learning approaches in the interpretation of endobronchial ultrasound images: a comparative analysis.

Abstract

This study explores the application of machine learning (ML) in analyzing endobronchial ultrasound (EBUS) images for the detection of lymph node (LN) malignancy, aiming to augment diagnostic accuracy and efficiency. We investigated whether ML could outperform conventional classification systems in identifying malignant involvement of LNs, based on eight established sonographic features. Retrospective data from two tertiary care hospital bronchoscopy units were utilized, encompassing healthcare reports of patients who had undergone EBUS between January 2017 and March 2023. The ML model was trained and tested using MATLAB, with 80% of the data allocated for training/validation, and 20% for testing. Performance was evaluated based on validation and testing accuracy, and receiver operating characteristic curves with comparing trained models and existing classification rules. The study analyzed 992 LNs, with 42.3% malignancy prevalence. Malignant LNs showed characteristic features such as larger size and distinct margins. The fine tuned models achieved testing accuracies of 95.9% and 96.4% for fine Gaussian SVM and KNN, respectively. Corresponding AUROC's were 0.955 and 0.963, outperforming other similar studies and conventional analyses. Fine tuned ML applications like SVM and KNN, can significantly enhance the analysis of EBUS images, improving diagnostic accuracy.