Clinical value of a deep learning multi-view fusion model for diagnosing fetal conotruncal defects
10.3760/cma.j.cn113903-20241207-00812
- VernacularTitle:超声多切面融合的深度学习识别模型诊断胎儿圆锥动脉干畸形的临床价值
- Author:
Hongmei GUO
1
;
Zhengxi DENG
;
Qiuhong XU
;
Sha WAN
;
Jianhua LUO
;
Shuangli REN
;
Shuxing ZHONG
;
Ting LEI
;
Xiaoyan MA
;
Yafui YAN
Author Information
1. 东莞市妇幼保健院超声科，东莞 523000
- Publication Type:Journal Article
- Keywords:
Fetal heart;
Congenital heart disease;
Conotruncal defect;
Multi-view fusion deep learning recognition model
- From:
Chinese Journal of Perinatal Medicine
2025;28(10):842-849
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop an ultrasound multi-view fusion recognition model and evaluate its clinical value in diagnosing fetal conotruncal defects (CTD).Methods:This prospective study collected cardiac ultrasound images from fetuses at 20-32 weeks of gestation undergoing prenatal ultrasound at Dongguan Maternal and Child Health Hospital between September 2022 and May 2024. The case group comprised fetuses diagnosed with CTD, while controls with normal cardiac structures were collected at a 1∶2 ratio. Both groups were divided into modeling training and validation sets at a 3∶1 ratio. One optimal standard image each from the four-chamber view, left ventricular outflow tract view, right ventricular outflow tract view, and three vessels and trachea view was included per fetus. A deep learning-based multi-view fusion recognition model was developed to differentiate normal conotruncal anatomy from CTD. Model performance was validated against post-abortion pathology or postnatal echocardiography results. SAS software was used for statistical analysis to calculate the sensitivity and specificity of three fusion models (based on positivity in any two, three, or four views, and were designated as Fusion Model 1, Fusion Model 2, and Fusion Model 3, respectively), with the optimal model determined by the maximum Youden index. Senior, intermediate, and junior prenatal sonologists independently diagnosed cases in the validation set under blinding conditions. Their diagnostic results were compared with those of the optimal model. Paired Chi-square test (Cochran's Q test) was employed to compare the differences between the diagnostic accuracy rates of sonologists at different experience levels and the sensitivity of the optimal model, thereby analyzing the auxiliary diagnostic value of the multi-view fusion recognition model. Results:The study included 88 CTD cases, excluding six cases (non-CTD diagnosed by post-abortion pathology or postnatal echocardiography or poor image quality), divided into 60 training and 22 validation cases (12 tetralogy of Fallot, four double outlet right ventricle, three transposition of great arteries, three persistent truncus arteriosus). The control group included 176 cases, excluding 15 cases (other cardiac abnormalities confirmed postnatally or poor image quality after re-evaluation), divided into 120 training and 41 validation cases. The sensitivities of Fusion Model 1, Fusion Model 2, and Fusion Mudel 3 were 0.86, 0.64, and 0.27, while their specificities were 0.76, 0.95, and 1.00, respectively. Fusion Model 1 demonstrated the highest Youden index (0.62) and was selected as optimal. Its diagnostic sensitivity showed no significant difference from senior sonologists [86% vs. 91% (20/22), Bonferroni-corrected P>0.999], but was significantly higher than intermediate [55% (12/22), Bonferroni-corrected P=0.049] and junior sonologists [32% (7/22), Bonferroni-corrected P=0.003]. Conclusion:The deep learning multi-view fusion model achieved diagnostic performance comparable to senior sonologists, demonstrating potential value in assisting CTD diagnosis, training less experienced sonologists, and supporting research and education.
