Using lightweight convolutional neural network to identify ventilation/perfusion scintigraphy for acute pulmonary...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Pulmonary embolism diagnostics

Background:

• Ventilation/perfusion (V/Q) scintigraphy interpretation is subjective and prone to errors.
• Objective quantification of V/Q scans is clinically challenging.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) architecture for objective V/Q quantification.
• To improve the accuracy and reduce subjectivity in V/Q scan analysis for pulmonary embolism (PE).

Main Methods:

• A lightweight CNN, PENet, was developed using depthwise separable convolution.
• PENet identifies defect areas in V/Q scans to calculate defect area percentage (DA%).
• Model performance was validated against clinical data from 288 patients with acute PE.

Main Results:

• PENet achieved high accuracy (87.47%), precision (89.22%), and F1-score (91.01%).
• Calculated DA% correlated significantly with clinical markers (NT-proBNP, age, troponin I, DLCO).
• PENet improved consistency between junior and senior radiologists (kappa=0.832).

Conclusions:

• PENet offers accurate and objective quantification of V/Q scans.
• Automated DA% calculation reflects PE severity and clinical data.
• PENet demonstrates significant potential for clinical application in PE diagnosis.