Development and validation of a deep learning model for multicategory pneumonia classification on chest computed...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate pneumonia diagnosis is crucial for effective management and reducing mortality.
• Chest CT scans can present overlapping features, complicating the differentiation of pneumonia from other conditions.
• Distinguishing between viral pneumonia, bacterial pneumonia, fungal pneumonia, and pulmonary tuberculosis is clinically significant.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for classifying five conditions on chest CT: viral pneumonia (VP), bacterial pneumonia (BP), fungal pneumonia (FP), pulmonary tuberculosis (PTB), and no pneumonia (NP).
• To evaluate the performance of the DL model against human radiologists.

Main Methods:

• A dataset of 1,776 chest CT cases was collected, with 1,611 used for model development and 165 for external validation.
• A 5-fold cross-validation approach was employed for model training.
• Radiologists' performance was assessed with and without the assistance of the DL model, using metrics like F1-score and AUC.

Main Results:

• The DL model achieved high F1-scores on the external test set: 0.933 for VP, 0.591 for BP, 0.848 for FP, 0.795 for PTB, and 0.976 for NP.
• The DL model demonstrated comparable performance across internal and external test sets.
• The DL model's F1-score surpassed that of radiologists, and DL assistance significantly improved radiologists' diagnostic F1-scores and precision.

Conclusions:

• The developed DL model effectively classifies various pneumonia types using chest CT imaging.
• The DL approach has the potential to enhance radiologists' diagnostic performance.
• Integrating DL tools into routine clinical workflows can improve pneumonia diagnosis and patient care.