PMFF-Net: A deep learning-based image classification model for UIP, NSIP, and OP.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• High-resolution computed tomography (HRCT) is crucial for diagnosing interstitial lung diseases (ILD), but its accuracy relies heavily on physician expertise.
• Differentiating common ILD types like usual interstitial pneumonia (UIP), non-specific interstitial pneumonia (NSIP), and organizing pneumonia (OP) can be challenging.

Purpose of the Study:

• To develop and evaluate a deep-learning-based classification model for differentiating common ILD types using HRCT.
• To provide a diagnostic reference tool to enhance physician accuracy in ILD diagnosis.

Main Methods:

• A dataset of 130 patient HRCT scans (UIP, NSIP, OP) and 50 normal scans was curated from four tertiary hospitals.
• The Parallel Multi-scale Feature Fusion Network (PMFF-Net) deep learning model was trained, validated, and tested.
• Model performance was assessed using accuracy, precision, recall, and F1-score, and compared against physician diagnoses.

Main Results:

• The PMFF-Net model achieved a diagnostic accuracy of 92.84% in classifying UIP, NSIP, OP, and normal imaging within 105 seconds for 18 images.
• The model's performance metrics (accuracy, precision, recall, F1-score) were all above 91%.
• Physician diagnostic accuracy varied significantly by experience and hospital level, with senior specialists outperforming junior doctors and internists.

Conclusions:

• The PMFF-Net model demonstrates high efficacy in classifying common ILD imaging types and normal scans.
• This AI tool can assist physicians across different hospital levels and departments in making timely and accurate clinical decisions for ILD diagnosis.