Simulating clinical features on chest radiographs for medical image exploration and CNN explainability using a...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Explainability of Convolutional Neural Networks (CNNs) is crucial for clinical adoption in radiology.
• Current attribution methods highlight image regions but don't explain underlying imaging features, hindering clinical use.
• A gap exists in understanding how CNNs interpret complex medical images and their clinical relevance.

Purpose of the Study:

• To introduce Semantic Exploration and Explainability using a Style-based Generative Adversarial Autoencoder Network (SEE-GAAN).
• To develop an AI explainability framework that semantically visualizes clinical and AI features in medical images.
• To enhance the interpretability of CNN predictions in radiological practice.

Main Methods:

• Developed SEE-GAAN, a novel framework utilizing latent space manipulation within a Style-based Generative Adversarial Autoencoder Network.
• Generated sequences of synthetic images to visualize feature manifestation and changes.
• Applied SEE-GAAN to a large cohort of chest radiographs (26,664 images, 15,409 patients) for clinical feature and AI prediction analysis.
• Used NT-pro B-type natriuretic peptide (BNPP) prediction as a proxy for acute heart failure to test the framework.

Main Results:

• SEE-GAAN sequences effectively visualized changes in anatomical and pathological morphology related to clinical and AI predictions.
• The framework clarified ambiguous regions identified by traditional attribution methods.
• Radiological interpretations confirmed SEE-GAAN's ability to capture clinically relevant imaging features.
• Demonstrated improved transparency of AI models compared to existing explainability techniques.

Conclusions:

• SEE-GAAN significantly enhances the explainability of CNNs in medical imaging.
• The framework aids in understanding clinical features for imaging biomarker discovery.
• SEE-GAAN improves AI transparency, facilitating its adoption in clinical radiology.
• This approach offers a more interpretable alternative to standard AI explainability methods.