Explainability of a Deep Learning Model for Mediastinal Lymph Node Station Classification in Endobronchial Ultrasound...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Accurate localization of thoracic lymph nodes is vital for lung cancer management.
• Deep learning (DL) models offer potential but often lack transparency.
• Explainable AI (XAI) tools like Grad-CAM can enhance DL model interpretability.

Purpose of the Study:

• To develop and evaluate a DL model for classifying thoracic lymph node stations using EBUS images.
• To quantitatively assess the anatomical relevance of Grad-CAM activations in EBUS imaging.
• To determine if XAI can provide meaningful insights into DL model behavior for EBUS.

Main Methods:

• A convolutional neural network (CNN) was trained on 35,527 labeled EBUS images for lymph node station classification.
• Grad-CAM was used to visualize model attention.
• Three expert bronchoscopists annotated Grad-CAM maps from 3,131 test images to assess anatomical relevance.

Main Results:

• The CNN achieved 63.1% accuracy in classifying lymph node stations, with high F1-scores for stations 4L, 4R, and 10R.
• Grad-CAM activations predominantly corresponded to lymph nodes and/or blood vessels.
• Expert annotation of Grad-CAM maps yielded 65.9% accuracy and 58.4% F1-score, with moderate interobserver agreement.

Conclusions:

• DL models can effectively assist in classifying thoracic lymph node stations during EBUS.
• XAI tools like Grad-CAM provide valuable insights into DL model decision-making processes.
• The proposed framework shows promise for enhancing anatomical orientation and training in EBUS procedures, warranting further validation.