Explainable emphysema detection on chest radiographs with deep learning.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Pulmonary emphysema diagnosis relies on identifying specific radiological signs on chest radiographs.
• Automated detection systems can aid in consistent and efficient diagnosis.

Purpose of the Study:

• To develop and evaluate a deep learning system for automatically detecting four explainable emphysema signs on frontal and lateral chest radiographs.
• To compare the performance of the explainable AI system against radiologists and a black-box AI model.

Main Methods:

• Retrospective collection of 3000 chest radiograph studies.
• Annotation of 4 key emphysema signs by two radiologists.
• Development of separate deep learning models for frontal and lateral images to predict visual signs and emphysema positivity.
• Performance evaluation using ROC and AUC on 422 held-out cases, with statistical comparisons (DeLong's and McNemar's tests).

Main Results:

• The system achieved high AUCs (0.924-0.946) for emphysema positivity prediction, comparable to radiologists.
• Performance was also comparable to a black-box model (AUCs 0.915-0.935).
• On cases with radiologist agreement, the model achieved an AUC of 0.981.

Conclusions:

• The proposed deep learning method accurately predicts emphysema positivity on chest radiographs.
• The system provides explainable visual sign labels, enhancing result interpretability.
• This approach offers performance on par with expert radiologists and existing black-box models.