Study of a deep learning-based artificial intelligence model for automatic measurement and classification of cystocele
10.3760/cma.j.cn131148-20241031-00563
- VernacularTitle:基于深度学习的人工智能模型在膀胱膨出自动测量与分型中的研究
- Author:
Ting XIAO
1
;
Xiduo LU
;
Yunqing CAO
;
Zhuoru LUO
;
Siyun DU
;
Yide QIU
;
Chaojiong ZHEN
;
Yinghong WEN
;
Dong NI
;
Weijun HUANG
Author Information
1. 佛山市第一人民医院超声诊疗中心，佛山 528000
- Publication Type:Journal Article
- Keywords:
Ultrasonography;
Artificial intelligence;
Pelvic floor;
Deep learning;
Dynamic video;
Cystocele
- From:
Chinese Journal of Ultrasonography
2025;34(4):334-339
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the clinical application value of convolutional neural network（CNN）based on deep learning in the automatic measurement of dynamic pelvic floor ultrasound video parameters and the diagnosis and classification of cystocele.Methods:A retrospective analysis was conducted on dynamic pelvic floor ultrasound videos from 398 postpartum women who underwent examinations at the First People's Hospital of Foshan between June 2020 and June 2022. The lowest point of the posterior bladder wall（PWB），urethral rotation angle（URA），and retrovesical angle（RVA）were manually measured by a senior radiologist（R1）and a junior radiologist（R2），and cystocele was classified according to the Green standard. The CNN model was employed to automatically extract the above parameters and to diagnose and classify cystocele. Using R1 measurements as a reference，intraclass correlation coefficient（ICC）was used to evaluate the consistency between the CNN model and R1，as well as between R2 and R1. The Kappa value was used to assess the agreement between the CNN model，R2，and R1 in the diagnosis and classification of cystocele. Additionally，the time consumption of the three measurement methods was compared.Results:The CNN model showed good consistency with R1 in measuring PWB and URA（ICC = 0.983，0.894），while its consistency in measuring RVA was moderate（ICC = 0.614）. The ICC between R2 and R1 in measuring PWB，URA，and RVA was 0.979，0.815，and 0.627，respectively. In the measurement of PWB and URA，the consistency between the CNN model and R1 was superior to that between R2 and R1. For cystocele diagnosis，the Kappa value between the CNN model and R1 was 0.924，which was higher than that between R2 and R1（0.904）. In cystocele classification，the Kappa value between the CNN model and R1 was 0.503，also higher than that between R2 and R1（0.426）. The CNN model processed a single video in 2.5（0.6）s，significantly faster than R1［59.9（16.9）s］and R2［56.8（11.2）s］（all P < 0.001）. Conclusions:The CNN model demonstrates high accuracy and efficiency in the measurement，diagnosis，and classification of cystocele，outperforming a junior radiologist and showing potential for clinical application.
