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

Background: Quantifying ventilation/perfusion (V/Q) scintigraphy and reducing errors caused by subjective interpretation of approximate grayscale images are clinically and analytically challenging.

Objectives: This study aims to objectively quantify V/Q results by developing a novel convolutional neural network architecture.

Methods: In this retrospective study, we collected data from patients with acute pulmonary embolism (PE). We proposed PENet, a lightweight neural network architecture based on depthwise separable convolution for identifying defect areas in V/Q scans. The defect area percentage (DA%) was obtained through threshold setting to quantify the mismatch range. The significance and accuracy of our model were verified by combining clinical data. We collected 4608 original scans from 288 patients as the preliminary dataset. We set the pixel threshold value to 30.

Results: PENet demonstrated accuracy (87.47 %), precision (89.22 %), and F1-score (91.01 %), superior to those of classical and other lightweight models. Spearman's rank correlation coefficient revealed correlations between DA% and N-terminal pro-brain natriuretic peptide, DA% and age, average DA% and age, average DA% and troponin I, DA% and age, and DA% and predicted percentage of diffusing lung capacity for carbon monoxide (P < .05). DA% (P = .004), DA% (P = .004), and average DA% (P = .006) differed significantly among PE risk groups. With the assistance of PENet, junior radiologists could achieve a high degree of consistency with senior radiologists (kappa=0.832, P < .001).

Conclusions: The accuracy of PENet reached 87.47 %. DA% calculated automatically could reflect PE severity and correlate well with clinical data. PENet shows promising results for clinical use.