Application of neural network model in ultrasound image segmentation of MTP1 tophus
10.3760/cma.j.cn131148-20250531-00298
- VernacularTitle:神经网络模型在MTP1痛风石超声图像分割中的应用
- Author:
Yuchen LI
1
;
Ting ZHANG
;
Yongming LIU
;
Lingtao WANG
;
Jiarui LIU
;
Yujie XIE
;
Cheng ZHAO
;
Jianrui DING
;
Chunping NING
Author Information
1. 青岛大学附属医院腹部超声科，青岛 266071
- Publication Type:Journal Article
- Keywords:
Tophus;
Ultrasonic diagnosis;
Automatic segmentation;
Neural network
- From:
Chinese Journal of Ultrasonography
2025;34(9):745-750
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To evaluate the performance of the neural network model in segmenting gout tophus in the first metatarsophalangeal（MTP1）joint ultrasound images.Methods:A total of 1 218 tophus images from 381 patients who underwent MTP1 ultrasound examinations in the Affiliated Hospital of Qingdao University between May 2023 and December 2024 were prospectively collected. The images were divided into training，validation，and test sets in a ratio of 7∶2∶1. Multiple neural network models were trained to automatically identify and segment tophus in the images，with physician-annotated tophus regions serving as the reference standard. Model performance was evaluated in the test set，and the impact of tophus characteristics（e.g.，echogenicity，size，and presence of bone erosion）on segmentation efficacy was analyzed.Results:In the test set，CMUNeXt demonstrated superior tophus segmentation performance versus Unet，Unet++，TransUnet，and CMU-Net，achieving an accuracy of 99.1%，precision of 79.1%，recall of 84.6%，intersection over union of 68.8%，and Dice similarity coefficient of 80.2%. Logistic regression identified tophus echogenicity，size，and bone erosion as independent efficacy factors OR（95% CI）=7.275（1.598-33.129），21.303（4.282-105.985），13.520（3.617-50.530），0.076（0.007-0.823）（all P<0.05）. Hypoechoic tophus demonstrated significantly superior segmentation performance compared to mixed-echoic and isoechoic tophus（all P<0.05），and lesions with larger maximum diameters（>10 mm）were segmented more effectively than smaller tophus（ P<0.05）. Conclusions:The CMUNeXt model enables accurate identification and segmentation of tophus in MTP1 ultrasound images，particularly excelling for larger and hypoechoic lesions. This approach holds significant promise for AI-assisted diagnosis of MTP1 gouty arthritis.
