Scholarlink

<p>Colorectal cancer (CRC) remains one of the leading causes of cancer-related deaths worldwide, necessitating accurate and efficient diagnostic tools. This study proposes an advanced deep learning-based framework for classifying CRC histological images into eight categories: Adipose, Complex, Debris, Empty, Lympho, Mucosa, Stroma, and Tumor. Initially, 18 pre-trained Convolutional Neural Networks (CNNs) were evaluated using transfer learning techniques. Statistical analysis via Duncan's Multirange test identified Darknet53 as the most effective model, achieving an accuracy (Acc.) of 92.43%, sensitivity (Sen.) of 92.43%, specificity (Spec.) of 98.51%, precision (Prec.) of 92.70%, and F1-score of 92.45%. To enhance the classification performance, Darknet53 was hybridized with a SVM by replacing the dense layer, and hyperparameters were optimized using a Random Grid Search algorithm. The optimized hybrid model exhibited a remarkable improvement, with an Acc. of 99.7%, Sen. of 99.7%, Spec. of 99.91%, Prec. of 99.98%, and F1-score of 99.98%, alongside significant improvements in other metrics. This hybrid approach demonstrates substantial potential in automating CRC diagnosis with high Acc., offering a valuable tool for clinical use.</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

Hybrid Darknet53-SVM model with random grid search optimization for enhanced colorectal cancer histological image classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月28日

查看全文
  • Pragati Patharia,
  • Prabira Kumar Sethy,
  • K. Lakshmipathi Raju,
  • Anita Khanna,
  • Ashoka Kumar Ratha,
  • Santi Kumari Behera,
  • Aziz Nanthaamornphong

摘要

Colorectal cancer (CRC) remains one of the leading causes of cancer-related deaths worldwide, necessitating accurate and efficient diagnostic tools. This study proposes an advanced deep learning-based framework for classifying CRC histological images into eight categories: Adipose, Complex, Debris, Empty, Lympho, Mucosa, Stroma, and Tumor. Initially, 18 pre-trained Convolutional Neural Networks (CNNs) were evaluated using transfer learning techniques. Statistical analysis via Duncan's Multirange test identified Darknet53 as the most effective model, achieving an accuracy (Acc.) of 92.43%, sensitivity (Sen.) of 92.43%, specificity (Spec.) of 98.51%, precision (Prec.) of 92.70%, and F1-score of 92.45%. To enhance the classification performance, Darknet53 was hybridized with a SVM by replacing the dense layer, and hyperparameters were optimized using a Random Grid Search algorithm. The optimized hybrid model exhibited a remarkable improvement, with an Acc. of 99.7%, Sen. of 99.7%, Spec. of 99.91%, Prec. of 99.98%, and F1-score of 99.98%, alongside significant improvements in other metrics. This hybrid approach demonstrates substantial potential in automating CRC diagnosis with high Acc., offering a valuable tool for clinical use.