Objective:To explore the accuracy and clinical application value of a Multi-Agent Reinforcement Learning framework (MARL framework) in three-dimensional ultrasound to automatically locate the coronal plane of the uterus.Methods:A total of 144 female patients who underwent routine gynecological examinations in Luohu People′s Hospital during May 2020 were selected as the experimental subjects. The three-dimensional volume data of the uterus of all the experimental subjects were collected by using the Resona-8 high-end color Doppler ultrasound system. A sonographer with more than 5 years of clinical experience manually locate the coronal plane of the uterus in all collected data, and at the same time automatically locate the coronal plane of the uterus MARL framework. The coronal plane images of the uterus obtained by the two methods were saved, and the operation time of the two methods was recorded. The coronal plane uterine images obtained by the two methods were mixed together, and the images were scored 0-1 by two senior ultrasound experts in a double-blind manner. The average score greater than or equal to 0.6 points was considered qualified.Results:①In 144 volunteers, among the coronal planes of the uterus located by the two methods, 131 were qualified by the manual method, and 137 were qualified by the automatic method.There was no statistical difference between the manual and automatic coronal plane images of the uterus (χ 2=1.934, P=0.164) by the chi-square test. ②Using interquartile range analysis, the median and interquartile range of the image score of the automatic group was 0.80(0.75, 0.90), while the median and interquartile range of the image score of the manual group was 0.80(0.75, 0.90). The Wilcoxon signed rank test was used to analyze the quality of the coronal plane images obtained by manual and automatic methods, and the difference was not statistically significant ( Z=1.241, P=0.215). ③The paired t test was used to compare the time required to locate the coronal surface of the uterus, by manual method (63.65±10.182)s, by automatic method (3.25±0.294)s, the difference between the two methods was statistically significant ( t=19.52, P<0.001). Conclusions:The method based on MARL framework has a high correlation with the manual locating of the coronal plane of uterus in three-dimensional ultrasound, and greatly reduces the operation time. It can be effectively applied in clinical practice and lays a foundation for the automatic diagnosis of uterine related diseases.

Objective:To explore the feasibility of extracting the key plane of the normal fetal palate on the 11-13 + 6 week from tomography ultrasonography imaging based on artificial intelligence. Methods:The fetal volume datas of 235 cases of 11-13 + 6 week normal fetal were collected from the Department of Ultrasound in the Luohu District People′s Hospital of Shenzhen and Huazhong University of Science and Technology Union Shenzhen Hospital from May 2020 to April 2021. The data acquisition was completed by sonographers A and B by using the GE Voluson E10 color Doppler ultrasound diagnostic instrument. All datas were marked offline by sonographer C. Tomographic imaging was performed on all included data by sonographer D, the tomographic images were saved and the time-consuming was recorded, and the datas of the sonographer group were obtained. The labeled data were randomly divided into the training set and test set for model transfer learning and testing.The 4-fold cross-validation was adopted to record the test set image output by the model and the time consumption to obtain the intelligent group data. A senior sonographer performed image analysis on the two groups of data images. The feasibility of the intelligent model was verified by comparing the score of the plane of retronasal triangle(RTP), the acquisition rate of RTP, the acquisition rate of the fault, and the time-consuming difference between the sonographer group and the intelligent group. Results:①There was no significant difference in the overall distribution of RTP scores between the sonographer group and intelligent group [5 (5, 6) points vs 5 (5, 6) points, Z=0.355, P=0.722]. The RTP acquisition rate of the sonographer group and intelligent group was not statistically significant (78.72% vs 76.60%, χ 2=0.55, P=0.458). The consistency and correlation of RTP obtained by the two groups were high (Kappa=0.645, φ=0.646, both P<0.001). ②The effective layers of the sonographer group were 9 (8, 9) and the intelligent group was 8 (7, 9). The fault acquisition rate of the doctor group was higher than that of the intelligent group (78.72% vs 68.51%, χ 2=12.52, P=0.001). The consistency and correlation of the two groups in obtaining faults were media (Kappa=0.503, φ=0.521, both P<0.001). ③The time-consuming of the intelligent group was significantly lower than that of the sonographer group [1.50 (1.23, 1.75)s vs 26.94 (22.28, 30.48)s, Z=11.440, P<0.001]. Conclusions:This research model can quickly and accurately realize the extraction and tomography of the key plane of the normal fetal palate on the 11-13 + 6 week.