Artificial intelligence-assisted quality control of anal sphincter ultrasound：a multicenter clinical study
10.3760/cma.j.cn131148-20250321-00152
- VernacularTitle:人工智能辅助肛门括约肌超声质控：一项多中心临床研究
- Author:
Man ZHANG
1
;
Junyan AN
;
Liang MU
;
Yuanchun FU
;
Kun WANG
;
Shuqing HUANG
;
Jiawei WU
;
Shuangyu WU
;
Ying CHEN
;
Ruixuan WANG
;
Xinling ZHANG
Author Information
1. 中山大学附属第三医院超声科，广州 510630
- Publication Type:Journal Article
- Keywords:
Ultrasonography;
Anal sphincter;
Quality control;
Artificial intelligence
- From:
Chinese Journal of Ultrasonography
2025;34(7):594-601
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a quality control model for anal sphincter ultrasound images and validate its diagnostic performance across multiple centers.Methods:A retrospective analysis was conducted on anal sphincter ultrasound images from seven medical centers in China between May 2019 and June 2022. A total of 7 040 images from 3 116 patients were included and divided into a training set（4 912 images）and a validation set（2 128 images）. The images were classified as standard or non-standard images by three experts. Three models were developed based on different image feature extraction methods：a single-branch model，a multi-branch weighted model，and a multi-branch ensemble model. The diagnostic performance of each model was evaluated using the area under the ROC curve（AUC），sensitivity，specificity，accuracy，positive predictive value，and negative predictive value，respectively. The optimal model was selected and compared with the performance of 4 doctors with varying experience levels. Sixty days later，the images with the assistance of the model's output were reassessed by the doctors to evaluate its impact on manual quality control.Results:① Among the 3 models，the multi-branch ensemble model demonstrated the highest AUC and sensitivity，with an AUC of 0.966（95% CI=0.958 - 0.974），a sensitivity of 91.83%，and a specificity of 91.41%. This model was named M quality. ② M quality's AUC was slightly lower than that of Senior A and B（0.966 vs. 0.976，0.976，and P<0.05），its sensitivity was slightly lower than that of Senior A（91.83% vs. 95.61%， P<0.001）but comparable to Senior B（91.83% vs. 92.89%， P=0.315），its specificity was slightly lower than Senior A and B（91.41% vs. 94.44%，98.18%，and P<0.05）. However，M quality significantly outperformed Junior A and B in AUC and sensitivity（AUC：0.966 vs. 0.850，0.818；sensitivity：91.83% vs. 84.90%，61.46%；all P<0.001），its specificity was higher than that of Junior A（91.41% vs. 80.28%， P<0.001）but lower than that of Junior B（91.41% vs. 95.96%， P<0.001）. ③ With model assistance，Senior B's sensitivity（92.89% vs. 94.20%， P=0.001）and Senior A's specificity（94.44% vs. 96.56%， P<0.001）improved significantly. Junior A and B showed significant improvements in AUC and sensitivity（AUC：0.931 vs. 0.850，0.914 vs. 0.818；sensitivity：91.83% vs. 84.90%，89.53% vs. 61.46%；all P<0.001）. After model assistance，Junior A's specificity increased（93.62% vs. 80.28%， P<0.001），while Junior B's specificity decreased（91.60% vs. 95.96%， P=0.013）. Conclusions:This study develops a quality control model for anal sphincter ultrasound images with robust diagnostic performance，approaching the level of seniors. The model significantly enhances the image quality assessment capabilities of juniors，demonstrating promising clinical application potential.
