Improving disease classification performance and explainability of deep learning models in radiology with heatmap...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiology and Diagnostic Tools
• Machine Learning in Healthcare

Background:

• Deep learning (AI) is increasingly used in radiology, necessitating explainability for clinical trust.
• Existing AI models require validation for diagnostic accuracy and transparency.
• Radiologist eye-gaze data offers insights into diagnostic decision-making processes.

Purpose of the Study:

• To improve disease classification performance in chest radiographs using a U-Net architecture.
• To enhance the explainability of AI models by generating heatmaps aligned with radiologist focus.
• To investigate the impact of incorporating heatmap generators and eye-gaze data into AI training.

Main Methods:

• Conducted three experiment sets using a U-Net architecture on a dataset of chest radiographs.
• Incorporated heatmap generators during training to enhance model focus visualization.
• Utilized multi-modal training with radiologist eye-gaze coordinates alongside image data.

Main Results:

• The best performing method achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.913 (95% CI [0.860, 0.966]).
• Significant improvements were observed in classifying 'pneumonia' (AUC 0.859) and 'congestive heart failure' (CHF) (AUC 0.962), addressing baseline model weaknesses.
• Generated heatmaps effectively highlighted key image regions for classification and showed improved alignment with radiologist eye-gaze data.

Conclusions:

• Integrating heatmap generators and eye-gaze information into AI training simultaneously improves disease classification accuracy.
• The proposed methods provide explainable AI visuals that correlate with expert radiologist attention.
• This approach enhances clinician trust and the utility of AI in radiological diagnosis.