深度学习在儿童心脏超声标准切面自动智能识别中的应用

Objective To explore the feasibility and accuracy of deep learning in automatic identification of standard views of pediatric echocardiography. Methods A total of 4 035 pediatric echocardiography images from the picture archiving and communication system database of Shanghai Children’s Medical Center, Shanghai Jiao Tong University School of Medicine, which were collected from Sep. to Oct. 2022, were selected and randomly divided into training set (2 421 images), validation set (807 images), and testing set (807 images) in a ratio of 6 ∶ 2 ∶ 2. A lightweight and efficient deep learning model was developed by improving DenseNet to achieve automatic identification of 15 standard views of pediatric echocardiography, and was compared with 3 commonly used deep learning models, including DenseNet121, InceptionV3, and MobileNetV3. With manual annotation results as the gold standard, the identification performance of deep learning models was evaluated using accuracy, precision, specificity, recall, and F1 score. The efficiency of the identification model was evaluated using 3 indicators: the number of parameters, model size, and floating-point operations. The identification results of the model were displayed using a confusion matrix, and the model’s concerns to image features were reflected using a heatmap. Results The average F1 scores of DenseNet121, InceptionV3, MobileNetV3, and the proposed model for identifying 15 standard and non-standard views of pediatric echocardiography were 94.59%, 95.13%, 92.41%, and 94.73%, the numbers of parameters were 7.0×10⁶, 24.4×10⁶, 4.2×10⁶, and 1.8×10⁶, the model sizes were 13.941, 48.777, 8.445, and 3.588 MB, and the floating-point operations were 11.16×10⁹, 12.89×10⁹, 0.86×10⁹, and 3.05×10⁹, respectively. The confusion matrix and heatmap showed that the proposed model had a higher recognition rate for 15 standard and nonstandard views of pediatric echocardiography than DenseNet121, InceptionV3 and MobileNetV3, and was able to focus on the key feature areas in the ultrasonic views. Conclusion The deep learning model proposed in this study can accurately identify standard cardiac ultrasound views in children; moreover, the model has a small number of parameters and can be operated with high efficiency.