Abstract
Lung cancer (LC) is the leading cause of cancer-related mortality worldwide. Endoscopic ultrasound-guided examination of thoracic lymph nodes is essential for appropriate treatment selection and patient prognosis. However, the procedure can be challenging technically due to complex anatomy, variable quality of ultrasound images, and considerable operator dependence. Artificial intelligence (AI) has the potential to improve anatomical interpretation, increase reproducibility, and support real-time clinical decision-making during endobronchial ultrasound (EBUS). This thesis investigates the use of deep learning (DL) in the interpretation of EBUS images during mediastinal LC staging.
The primary objective of this doctoral project was to develop and evaluate DL-based approaches for automated analysis of EBUS images, focusing on three tasks: segmentation of anatomical structures, classification of mediastinal lymph node stations, and structured assessment of model explainability.
Three prospective studies were conducted using EBUS data acquired during routine EBUS procedures. In the first study, a U-Net–based DL model was developed to automatically segment mediastinal lymph nodes and vascular structures. This study demonstrated strong model performance with regards to detection, and processing speeds suitable for real-time applications in an intraoperative setting.
The second study presented a DL–based method for automated classification of mediastinal lymph node stations from sequences of EBUS images. Although overall classification accuracy of this model was moderate, the prediction speed was close to real-time, providing proof-of-concept for the selected approach.
The third study evaluated the interpretability of the lymph node station classification model using Gradient-weighted Class Activation Mapping (Grad-CAM). Expert assessments showed that the model’s attention consistently aligned with clinically relevant anatomical landmarks, such as lymph nodes and vascular structures. This finding supports the clinical plausibility of the model’s predictions and marks an important step toward more transparent DL-assisted interpretation of EBUS images.
Collectively, these studies provide proof-of-concept for intraoperative DL analysis in EBUS imaging using prospectively collected clinical data. The results suggest that DL-assisted analysis may support anatomical orientation and increase transparency in the assessment of EBUS images. Future work is required before clinical implementation of the proposed methods, including multicenter validation, larger datasets, and a systematic evaluation of how they can be integrated into the clinical workflow.