Radiologist-Trained AI Model for Identifying Suboptimal Chest-Radiographs.

Area of Science:

• Radiology and Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Suboptimal chest radiographs (CXRs) can hinder the interpretation of critical findings.
• Accurate image acquisition is crucial for effective diagnosis in radiology.

Purpose of the Study:

• To evaluate radiologist-trained AI models for differentiating between suboptimal (sCXR) and optimal (oCXR) chest radiographs.
• To assess the performance of AI in identifying specific causes of CXR suboptimality.

Main Methods:

• A retrospective study of 3278 adult CXRs from 5 sites.
• AI models were trained on 2202 CXRs and tested on 1076 CXRs.
• Performance was analyzed using Area Under the Curve (AUC) for classification accuracy.

Main Results:

• AI models demonstrated high accuracy in classifying oCXR and sCXR, with AUCs ranging from 0.87 to 0.94 for various suboptimality causes.
• Specific findings like obscured anatomy, inadequate exposure, low lung volume, and patient rotation were identified with high sensitivity and specificity.
• AI achieved 91% accuracy for missing anatomy and 95% accuracy for patient rotation.

Conclusions:

• Radiologist-trained AI models can accurately classify optimal and suboptimal CXRs.
• Implementing these AI models at the equipment's front end can assist radiographers in repeating suboptimal studies.
• This AI application has the potential to improve the quality of chest radiography and diagnostic reliability.