Prototype early diagnostic model for invasive pulmonary aspergillosis based on deep learning and big data training

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Current invasive pulmonary aspergillosis (IPA) diagnosis relies on integrated clinical, radiological, and microbiological data.
• Limited research exists on applying artificial intelligence (AI) to IPA diagnosis.
• There is a need for non-invasive, objective diagnostic tools for IPA.

Purpose of the Study:

• To develop a non-invasive, objective, and user-friendly AI approach for early IPA diagnosis.
• To create a deep learning model for automated IPA detection.
• To improve diagnostic accuracy and speed for IPA.

Main Methods:

• A prototype deep learning model, IPA-NET, was developed using transfer learning on 300,000 non-fungal pneumonia CT images.
• IPA-NET was trained and internally validated using clinical features and chest CT images from early-stage IPA and non-fungal pneumonia patients.
• The model's performance was further assessed using an independent external test set.

Main Results:

• IPA-NET achieved high diagnostic performance on the internal test set: 96.8% accuracy, 0.98 sensitivity, 0.96 specificity, and 0.99 AUC.
• Validation on the external test set showed IPA-NET maintained strong performance: 89.7% accuracy, 0.88 sensitivity, 0.91 specificity, and 0.95 AUC.
• The model demonstrated reliable diagnostic capabilities across different datasets.

Conclusions:

• The developed deep learning model, IPA-NET, offers a novel, non-invasive, and objective method for early IPA diagnosis.
• IPA-NET shows significant potential to aid clinicians in timely and accurate IPA detection.
• This AI-driven approach represents a promising advancement in diagnosing invasive pulmonary aspergillosis.