Scholarlink

<p>Chronic liver disease (CLD) segmentation and classification present significant challenges due to inaccurate diagnosis and misleading clinical treatment processes. In contemporary times, artificial intelligence (AI) and deep learning (DL) models are considered important tools for driving quantitative biomarkers for efficient stratification and computer-aided diagnosis support. However, prevailing DL models necessitate further enhancements as they rely on flawed assumptions that tumors exhibit non-complex physical structures and uniform image boundaries. To resolve this concern, this paper suggests a novel dual encoder deep learning structure named DEDSWIN-Net to mitigate this problem. The proposed framework consists of four components: a dilated convolution-based encoder, a transformer-based encoder, a multi-scale multiple feature fusion decoder (MSMFD), and a DL training model. The dilated convolution layers can retrieve fine spatial features, whereas Swin models are designed to obtain global information. Both these features are integrated by the MSMFD component that enhances segmentation. Finally, the deep learning training model is constructed using feed-forward principles for better classification of CLD. Comprehensive evaluation is conducted utilizing the LiTS2017-CT Image dataset comprising 100 hepatic disorders, and its performance is evaluated using ten powerful metrics: Jaccard Index (Jcc), dice similarity coefficient (DICE), precision (P), accuracy (Acc), specificity (Spec), recall (R), F1 score (FS), average symmetric surface distance (ASSD), Hausdorff distance (HD), and Intersection over Union (IoU). Furthermore, existing state-of-the-art DL frameworks are considered for comprehensive examination. Results demonstrate that the suggested methodology reveals supreme performance in terms of segmentation and classification with DICE: 0.984, JC: 0.92, IoU: 0.02, precision: 0.95, recall: 0.940, ASSD: 0.64, and HD: 0.32. Moreover, the exploratory outcomes prove the effectiveness of the suggested methodology in obtaining relevant characteristics, fuelling a better computer-aided diagnosis system for CLD.</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

DEDSWIN-Net: Dual Encoder Dilated Convolution and Swin Transformer Network for the Classification of Liver CT Images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月30日

查看全文
  • Jyoshna Allenki,
  • Hemant Kumar Soni

摘要

Chronic liver disease (CLD) segmentation and classification present significant challenges due to inaccurate diagnosis and misleading clinical treatment processes. In contemporary times, artificial intelligence (AI) and deep learning (DL) models are considered important tools for driving quantitative biomarkers for efficient stratification and computer-aided diagnosis support. However, prevailing DL models necessitate further enhancements as they rely on flawed assumptions that tumors exhibit non-complex physical structures and uniform image boundaries. To resolve this concern, this paper suggests a novel dual encoder deep learning structure named DEDSWIN-Net to mitigate this problem. The proposed framework consists of four components: a dilated convolution-based encoder, a transformer-based encoder, a multi-scale multiple feature fusion decoder (MSMFD), and a DL training model. The dilated convolution layers can retrieve fine spatial features, whereas Swin models are designed to obtain global information. Both these features are integrated by the MSMFD component that enhances segmentation. Finally, the deep learning training model is constructed using feed-forward principles for better classification of CLD. Comprehensive evaluation is conducted utilizing the LiTS2017-CT Image dataset comprising 100 hepatic disorders, and its performance is evaluated using ten powerful metrics: Jaccard Index (Jcc), dice similarity coefficient (DICE), precision (P), accuracy (Acc), specificity (Spec), recall (R), F1 score (FS), average symmetric surface distance (ASSD), Hausdorff distance (HD), and Intersection over Union (IoU). Furthermore, existing state-of-the-art DL frameworks are considered for comprehensive examination. Results demonstrate that the suggested methodology reveals supreme performance in terms of segmentation and classification with DICE: 0.984, JC: 0.92, IoU: 0.02, precision: 0.95, recall: 0.940, ASSD: 0.64, and HD: 0.32. Moreover, the exploratory outcomes prove the effectiveness of the suggested methodology in obtaining relevant characteristics, fuelling a better computer-aided diagnosis system for CLD.