Scholarlink

The use of Wireless Capsule Endoscopy (WCE) has changed gastrointestinal diagnostics completely by allowing us to image the entire digestive system noninvasively. In addition to taking a great deal of time, manually analyzing WCE images for bleeding is prone to human error. In this paper, we present a unique deep learning architecture that combines the advantages of Kolmogorov-Arnold Networks (KANs) and Convolutional Neural Networks (CNNs) for automatically detecting bleeding in WCE images. A CNN backbone consists of four convolutional blocks that hierarchically extracts high-level features from raw WCE images. By using the Kolmogorov-Arnold representation theorem, the KAN layers describe complicated, nonlinear relationships using adaptive B-spline transformations. KAN layers use an entropy-based regularization term as well as a dual-pathway mechanism that integrates linear and nonlinear transformations to avoid overfitting. Furthermore, the KAN layers’ adaptive grid architecture dynamically adapts to the distribution of input features, making the model more capable of detecting minute bleeding patterns. According to experimental results on a benchmark WCE dataset, the proposed architecture outperforms conventional CNN-based methods in bleeding classification. Through the integration of CNNs with KANs, this work paves the way for more accurate and effective WCE diagnostics.

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

Adaptive Spline Transformations vs Traditional Linear Layers: A Comparative Analysis of CNN Architectures for WCE Images Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yasmina El Khalfaoui,
  • Chaima Elmejgari,
  • Brahim Alibouch,
  • Younes Nadir,
  • Ahmed Fouad El Ouafdi

摘要

The use of Wireless Capsule Endoscopy (WCE) has changed gastrointestinal diagnostics completely by allowing us to image the entire digestive system noninvasively. In addition to taking a great deal of time, manually analyzing WCE images for bleeding is prone to human error. In this paper, we present a unique deep learning architecture that combines the advantages of Kolmogorov-Arnold Networks (KANs) and Convolutional Neural Networks (CNNs) for automatically detecting bleeding in WCE images. A CNN backbone consists of four convolutional blocks that hierarchically extracts high-level features from raw WCE images. By using the Kolmogorov-Arnold representation theorem, the KAN layers describe complicated, nonlinear relationships using adaptive B-spline transformations. KAN layers use an entropy-based regularization term as well as a dual-pathway mechanism that integrates linear and nonlinear transformations to avoid overfitting. Furthermore, the KAN layers’ adaptive grid architecture dynamically adapts to the distribution of input features, making the model more capable of detecting minute bleeding patterns. According to experimental results on a benchmark WCE dataset, the proposed architecture outperforms conventional CNN-based methods in bleeding classification. Through the integration of CNNs with KANs, this work paves the way for more accurate and effective WCE diagnostics.