Neural network analysis of ventilation-perfusion lung scans

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Nuclear Medicine
• Cardiopulmonary Diagnostics

Background:

• Interpretation of ventilation-perfusion (V/Q) lung scans is crucial for diagnosing pulmonary embolism.
• Current interpretation algorithms can be complex and may benefit from advanced analytical methods.
• The need for objective and accurate diagnostic tools in pulmonary embolism assessment is high.

Purpose of the Study:

• To develop and evaluate a neural network model for interpreting V/Q lung scans.
• To compare the diagnostic performance of the neural network against experienced human observers.
• To assess the potential of artificial intelligence in improving pulmonary embolism diagnosis.

Main Methods:

• A neural network was designed with 28 input parameters derived from V/Q scan findings.
• The network utilized a single hidden layer with 10-20 nodes and outputted probability of pulmonary embolism.
• Model training involved 100 V/Q scans with pulmonary angiographic correlation; testing used 28 new scans.

Main Results:

• The neural network, specifically with 15 hidden nodes, demonstrated superior performance compared to an experienced observer.
• The network's predictions for pulmonary embolism likelihood were statistically more accurate in the test set (P = .039).
• The AI model successfully synthesized multiple V/Q findings into a single diagnostic probability.

Conclusions:

• Neural networks offer a powerful tool for interpreting complex V/Q lung scan data.
• This AI model shows potential to enhance diagnostic accuracy for pulmonary embolism.
• The ability of neural networks to learn and adapt presents advantages over existing interpretation algorithms.