Liver ultrasound image classification model based on bimodal fusion and deep residual network
10.19745/j.1003-8868.2025175
- VernacularTitle:基于双模态融合与深度残差网络的肝脏超声图像分类模型研究
- Author:
Xiao-yan YE
1
;
Xing SU
;
Xiao-lin LI
;
Le-yu ZHANG
Author Information
1. 广州科技贸易职业学院生物技术与健康学院,广州 511400
- Publication Type:Journal Article
- Keywords:
liver ultrasound;
ultrasonography;
space-occupying lesion of liver;
bimodal fusion;
image classification;
ResNet-18 network;
deep residual network
- From:
Chinese Medical Equipment Journal
2025;46(10):9-16
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a liver ultraound image classification model based on bimodal fusion and deep residual network to enhance the diagnosis accuracy of space-occupying lesions(SOLs)of liver.Methods Firstly,a liver ultrasound dataset containing 164 lesions from 100 patients was constructed,including the ultrasound and ultrasonography videos of the patients.Secondly,the parallel residual block was introduced to improve the ResNet-18 network,and the simple attention module and coordinate attention mechanism were used to extract the image features of ultrasound and echography videos,respectively.Finally,a bimodal fusion network was developed with the image features of ultrasound and sonography videos,which was combined with the improved ResNet-18 network to form an ultrasound image classification model.The ultrasound image classification model proposed underwent performance verification by ablation experiment,application evaluation,diagnosis efficacy evaluation,significance evaluation for assisting surgical operation and comparison with the existing classification models in terms of image classification ability.Results The ablation experiment results showed that the proposed model performed the best in classification speed and accuracy when compared with the existing classification models,with a floating-point operation speed of 5.328×109/s,an average accuracy of 0.941 and a calculation speed of 245.266 frames/s.The application evaluation results indicated when compared with the existing classification models the proposed model had the best convergence performance of the loss function curve and the lowest misdiagnosis rate of 7.03%.The diagnosis efficacy evaluation results proved the proposed model gained advantages over other models in diagnostic efficacy,with a sensitivity of 89.38%,a specificity of 94.12%,an accuracy of 90.85%,a positive predictive value of 97.12%and a negative predictive value of 80.00%.The significance evaluation for assisting surgical operation found when compared with the existing classification models the proposed model had the shortest end-to-end delay of 723 ms;laparoscopic hepatectomy assisted with the proposed model had the blood loss reduced by 109.832 mL when compared with the traditional laparoscopic procedure,with the difference being statistically significant(P<0.05).Conclusion The proposed model enhances the diagnosis efficiency and accuracy of ultrasond SOLs of liver,providing support for clinical diagnosis and surgical assistance.[Chinese Medical Equipment Journal,2025,46(10):9-16]
