To propose a deep learning-based view recognition model, SlowFast-Echo, for the automatic view recognition of two-dimensional (2D) transthoracic echocardiographic dynamic images.

From August to December 2022, 722 patients who underwent 2D transthoracic echocardiography at the Department of Ultrasound Medicine, Affiliated Hospital of Medical School, Nanjing University (9 types of clinically commonly used views [including apical two-chamber, apical three-chamber, and apical four-chamber views], with a total of 2243 dynamic images) were selected, and the images of each view were divided into training set, validation set, and test set in a ratio of 5:2:3. After training and validation of the SlowFast-Echo model, the performance of the model was evaluated quantitatively in terms of accuracy, precision, recall, and F1 score, qualitatively in terms of the interpretability of the model with regard to class activation mapping, and practically in terms of the performance of the model after field deployment to the ultrasound medicine department.

The overall accuracy, precision, recall, and F1 score of the SlowFast-Echo model for dynamic image view recognition in the test set were 0.9866, 0.9847, 0.9872, and 0.9859, respectively, and the significance heatmap indicated that the model's regions of interest were generally consistent with those drawn by the physicians; e.g., the model accurately pinpointed the significant aorta and aortic valve in parasternal short axis view of great vessel (PSAX_GV) view, mitral valve in parasternal short axis view of left ventricle at mitral value level (PSAX_MV) view, and papillary muscles in parasternal short axis view of left ventricle at papillary muscle level (PSAX_PM) view. The overall accuracy, precision, recall, and F1 score of the model for view recognition after deployment were 0.9903, 0.9865, 0.9868, and 0.9865, respectively, and the average inference time on RTX 3060 GPU for a single dynamic image was (303.2±119.3) ms, which basically meets the clinical demand for immediate processing after image acquisition.

The SlowFast-Echo model proposed in this study has good performance in view recognition of 2D transthoracic echocardiographic dynamic images and inference in real time, which is practically useful.