Developing Echocardiography Image Quality Management System Based on Deep Learning

Overview

To develop an echocardiography image quality management system based on deep learning to achieve objective and accurate automatic echocardiography image quality control. A total of 2000 patients performing transthoracic echocardiography were prospectively enrolled in the Department of Ultrasound Medicine of the Affiliated Drum Tower Hospital with Medical School of Nanjing University. The data of 8 TTE view segmentations were collected, including the views of the parasternal long axis of the left ventricle (PLAX_LV), parasternal short axis of the large vessel level (PSAX_GV), parasternal short axis of the mitral valve level (PSAX_MV), parasternal short axis of the papillary muscle level (PSAX_PM), parasternal short axis of the apical level (PSAX_AP), apical four cavity (A4C), apical three cavity (A3C), apical two cavity (A2C). The data of 1500 patients were used as the training set, and the rest were used as the validation set. These video data were classified into corresponding view segmentations and analyzed by the Video Swin Transformed Model. Then, the scoring module of different view segmentations combined key frame extraction, image segmentation, video target recognition and video classification model were established. At the same time, the scores achieved by the automatic echocardiography image assessment system were compared with the artificial score. By constantly correcting and learning and eventually building an primary automated grading system. At last, the automatic echocardiography image assessment system was constructed and performed on the rest 500 patients.

Description

To develop an echocardiography image quality management system based on deep learning to achieve objective and accurate automatic echocardiography image quality control. A total of 2000 patients performing transthoracic echocardiography were prospectively enrolled in the Department of Ultrasound Medicine of the Affiliated Drum Tower Hospital with Medical School of Nanjing University. The inclusion criteria: Patients with standardized TTE view segmentation; The exclusion criteria: Patients with incomplete standard segmentations. The data of 8 TTE view segmentations were collected, including the views of the parasternal long axis of the left ventricle (PLAX_LV), parasternal short axis of the large vessel level (PSAX_GV), parasternal short axis of the mitral valve level (PSAX_MV), parasternal short axis of the papillary muscle level (PSAX_PM), parasternal short axis of the apical level (PSAX_AP), apical four cavity (A4C), apical three cavity (A3C), apical two cavity (A2C). The data of 1500 patients were used as the training set, and the rest were used as the validation set. These video data were classified into corresponding view segmentations and analyzed by the Video Swin Transformed Model. Then, the scoring module of different view segmentations combined key frame extraction, image segmentation, video target recognition and video classification model were established. At the same time, the scores achieved by the automatic echocardiography image assessment system were compared with the artificial score. By constantly correcting and learning and eventually building an primary automated grading system. At last, the echocardiography image quality management system was performed on the rest 500 patients and improved.

Eligibility

Inclusion Criteria:

1. aged ≥18years, gender unlimited;
2. Patients with standardized TTE views;
3. Subjects participated in the study voluntarily and signed informed consent;

Exclusion Criteria:

1. patients wirh incomplete standard TTE views;
2. patients with poor sound transmission conditions.