Collaborative Integration of AI and Human Expertise to Improve Detection of Chest Radiograph Abnormalities

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Computer-Aided Diagnosis

Background:

• Perceptual errors in chest radiograph interpretation can lead to misdiagnosis.
• Integrating multimodal data, such as eye gaze and reports, holds potential for improving diagnostic accuracy.
• Existing AI systems may not fully address the collaborative aspect of radiological diagnosis.

Purpose of the Study:

• To develop a collaborative artificial intelligence (AI) system, named CoRaX, for chest radiograph interpretation.
• To improve diagnostic accuracy by integrating eye gaze data and radiology reports to identify and correct perceptual errors.
• To evaluate the system's performance in collaborative diagnostic settings.

Main Methods:

• Developed CoRaX, a collaborative AI solution using a large multimodal model.
• Utilized public datasets REFLACX and EGD-CXR for training and development.
• Evaluated the system on simulated error datasets with random and uncertainty-based alterations, focusing on referral quality and collaborative performance.

Main Results:

• CoRaX corrected 21.3% of errors in the random masking dataset and 34.6% in the uncertainty masking dataset.
• The system achieved 63.0% accuracy in identifying regions of missed abnormalities (random masking) and 58.0% (uncertainty masking).
• Satisfactory interactions between CoRaX and radiologists were high, at 85.7% and 78.4% for the respective datasets, indicating effective diagnostic aid.

Conclusions:

• The CoRaX system effectively collaborates with radiologists to address perceptual errors in chest radiographs.
• The integration of eye gaze data and radiology reports enhances the AI's ability to rectify diagnostic mistakes.
• CoRaX shows promise for improving the reliability and accuracy of chest radiograph interpretation in clinical settings.