Scholarlink

<p>Liver Fibrosis (LF) is a progressive liver disease that requires accurate staging for effective treatment. The conventional manual diagnostic methods are more time-consuming, labor-intensive, and susceptible to human error. Additionally, medical image datasets often contain noise and poor quality, which degrade Deep Learning (DL) model performance. Meanwhile, conventional aggregation operators often fail to fully leverage the unique strengths of individual DL models by highlighting the need for more robust aggregation methods. To address these challenges, we propose a novel ensemble network based on Dempster-Shafer theory-driven fusion (DSFuse-Net). The proposed approach begins with an advanced preprocessing phase where liver ultrasound images undergo contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) and noise suppression via Non-Local Means filtering. These enhanced images are then fused using Daubechies wavelet transform by leveraging both low- frequency and high-frequency decompositions to improve image quality and highlight critical fibrosis features. For robust feature extraction, we carefully selected three pretrained DL architectures including DenseNet201, DenseNet169, and ResNet50V2 for considering them as base models. These models serve as candidates for fusion using Dempster-Shafer theory (DST), which enables dynamically weights predictions based on confidence. The proposed approach effectively integrates evidence by handling uncertainty and conflicting information, thereby improving the overall decision-making process. We validated DSFuse-Net on a publicly available LF dataset and achieved superior accuracy of 98.42% among five LF stages labelled as F0 to F4. We evaluate model performance using key metrics and analyze results with confusion-matrices, Receiver-Operating-Characteristic (ROC) curves, and Precision-Recall (PR) curves to assess classification accuracy. Additionally, Grad-CAMs and SHAP (SHapley Additive exPlanations) are employed to provide visual explanations and interpret feature contributions. Furthermore, statistical significance is validated using McNemar’s and the Wilcoxon Signed-Rank test to confirm that the findings are not due to random chance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DSFuse-Net: A Novel Uncertainty-Aware Dynamic Dempster-Shafer Theory-Guided Ensemble Network for Liver Fibrosis Grading

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Omair Bilal,
  • Sohaib Asif

摘要

Liver Fibrosis (LF) is a progressive liver disease that requires accurate staging for effective treatment. The conventional manual diagnostic methods are more time-consuming, labor-intensive, and susceptible to human error. Additionally, medical image datasets often contain noise and poor quality, which degrade Deep Learning (DL) model performance. Meanwhile, conventional aggregation operators often fail to fully leverage the unique strengths of individual DL models by highlighting the need for more robust aggregation methods. To address these challenges, we propose a novel ensemble network based on Dempster-Shafer theory-driven fusion (DSFuse-Net). The proposed approach begins with an advanced preprocessing phase where liver ultrasound images undergo contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) and noise suppression via Non-Local Means filtering. These enhanced images are then fused using Daubechies wavelet transform by leveraging both low- frequency and high-frequency decompositions to improve image quality and highlight critical fibrosis features. For robust feature extraction, we carefully selected three pretrained DL architectures including DenseNet201, DenseNet169, and ResNet50V2 for considering them as base models. These models serve as candidates for fusion using Dempster-Shafer theory (DST), which enables dynamically weights predictions based on confidence. The proposed approach effectively integrates evidence by handling uncertainty and conflicting information, thereby improving the overall decision-making process. We validated DSFuse-Net on a publicly available LF dataset and achieved superior accuracy of 98.42% among five LF stages labelled as F0 to F4. We evaluate model performance using key metrics and analyze results with confusion-matrices, Receiver-Operating-Characteristic (ROC) curves, and Precision-Recall (PR) curves to assess classification accuracy. Additionally, Grad-CAMs and SHAP (SHapley Additive exPlanations) are employed to provide visual explanations and interpret feature contributions. Furthermore, statistical significance is validated using McNemar’s and the Wilcoxon Signed-Rank test to confirm that the findings are not due to random chance.