Neural networks in ventilation-perfusion imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pulmonary embolism (PE) diagnosis relies on imaging techniques.
• Ventilation-perfusion (V-P) scans are crucial for PE assessment.
• Optimizing diagnostic accuracy in V-P scan interpretation is essential.

Purpose of the Study:

• To enhance artificial neural network (ANN) performance for PE prediction using V-P scans.
• To identify optimal V-P scan features and network parameters for accurate PE diagnosis.

Main Methods:

• ANNs were developed incorporating V-P scan criteria like defect sharpness and perfusion completeness.
• Quantification of abnormalities used a continuous numeric scale.
• Network parameters and training iterations (150 training, 30 testing cases) were systematically optimized and compared to pulmonary angiography.

Main Results:

• ANNs achieved performance comparable to experienced nuclear medicine physicians with minimal V-P scan features.
• Optimal network training involved 50-100 iterations; longer training reduced performance.
• Limited, unconventional V-P scan features proved effective for PE prediction.

Conclusions:

• Effective ANNs for PE prediction can be built using a select set of V-P scan features.
• Network performance is sensitive to parameter tuning and training duration.
• This approach offers a promising avenue for improving V-P scan interpretation in PE diagnosis.