Ultrasound image intelligent diagnosis in community-acquired pneumonia of children using convolutional neural network based transfer learning (Preprint)

Abstract

Studies show that lung ultrasound (LUS) can accurately diagnose pneumonia in children, but intelligent diagnosis hasn’t been explored in this area.

To construct deep learning (DL) models based on transfer learning (TL) to explore the feasibility of ultrasound image diagnosis and grading in community-acquired pneumonia (CAP) of children.

From September 2021 to February 2022, 89 inpatients who were expected to receive a diagnosis of CAP in the pediatric ward of local hospital were prospectively enrolled. Clinical data were collected, a LUS images database was established, and the diagnostic values of LUS in CAP were analyzed. We constructed DL models using AlexNet, ResNet-18 and ResNet-50 to perform CAP diagnosis and grading on the LUS database and evaluated the performance of each model. The models were trained separately with transfer learning.

1. Among the 89 children, 24 were in the non-CAP group, and 65 were finally diagnosed with CAP, including 44 in the mild group and 21 in the severe group. 2. LUS was highly consistent with clinical diagnosis, CXR and chest CT (kappa values = 0.943, 0.837, 0.835). 3. In the task of diagnosing CAP in children, different ratios of training and test sets (5:5; 8:2; 9:1) affected the performance of the model. The accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and area under the receiver operating characteristic curve of the AlexNet model ranged from 87.6%–89.5%, 93.2%–97.1%, 68.4%–73.6%, 89.5%–90.7%, 79.5%–89.9%, and 0.820–0.841; for the ResNet-18 model, these values were 87.3%–92.4%, 94.4%–98.3%, 64.6%–76.1%, 89.4%–91.8%, 83.7%–94.9%, and 0.795–0.972; moreover, for the ResNet-50 model, these values were 88.2%–90.9%, 94.0%–97.9%, 70.8%–72.1%, 90.2%–91.0%, 81.3%–92.6%, and 0.824–0.850. 4. When the training set and test set ratio was 8:2, the AlexNet, ResNet-18, and ResNet-50 models were highly consistent with the manual diagnosis CAP (kappa values = 0.832, 0.848, and 0.847 respectively), which was comparable to CXR and chest CT, and the ResNet-18 model performed better than manual ultrasound diagnosis (P=0.021). 5. In the task of grading, the accuracy of all three models increased with additions to the training set. When the ratio was 9:1, the accuracy of the ResNet-18 model reached 96%.

LUS is a reliable method for diagnosing CAP in children. The transfer learning-based DL models AlexNet, ResNet-18 and ResNet-50 perform well in children’s CAP diagnosis in the database we established; of these, the ResNet-18 model achieves the best performance and may serve as a tool for the further research and development of AI automatic diagnosis LUS system in clinical applications.

This study was a prospective case-control clinical diagnostic study, which was reviewed by the clinical trial registration website www.chictr.org.cn and obtained the registration number ChiCTR2200057328.