Scholarlink

<p>Lung cancer remains the second leading cause of death globally, with a significantly higher mortality rate compared to other cancer types. In recent years, deep learning has emerged as a transformative approach in medical image analysis, offering promising capabilities for enhancing cancer detection and diagnosis. Despite its potential, deep learning models often suffer from overfitting, which can substantially degrade their performance, especially when data variability is limited. To address these challenges, this study proposes a novel and comprehensive framework for lung cancer classification using computed tomography (CT) images, which involves several key stages. The framework begins with a data augmentation, applying techniques such as rotation, flipping, and scaling to improve data diversity and model robustness. Preprocessing is then carried out using median filtering to suppress image noise and enhance quality for subsequent analysis. Subsequently, an Improved SegNet (ISN) model is introduced for segmentation, incorporating a Mixed Average Lp (MALp) pooling layer. This proposed pooling mechanism enhances the retention of spatial information and improves the segmentation accuracy and efficiency. Then, pertinent features such as Modified Pyramid Histogram of Oriented Gradients (MPHOG), color and shape descriptors such as epsilon, area, perimeter, and hull are obtained from the segmented outcome, along with deep features extracted using the ResNet, VGG-16, and GoogLeNet. The obtained feature set is subjected to the classification phase, where a hybrid deep learning model is proposed for classifying lung cancer that integrates an Improved LeNet with LinkNet models and produces the final classified outcomes as Normal, Benign, and Malignant. Moreover, the experimental results demonstrate the superiority of the proposed framework over traditional methods, achieving notable performance metrics including an accuracy of 0.981, precision of 0.975, and <i>F</i>-measure of 0.942. These results underscore the practical significance of the framework in supporting early and accurate lung cancer diagnosis, ultimately contributing to improved clinical outcomes and potentially saving the lives of patients.</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

Improved LeNet–LinkNet Architecture for Lung Cancer Classification with Improved Segnet Architecture and Deep Features

  • 期刊论文

  • 发表日期：

    2025年10月19日

查看全文
  • D. Nagaraju,
  • B. P. Santosh Kumar

摘要

Lung cancer remains the second leading cause of death globally, with a significantly higher mortality rate compared to other cancer types. In recent years, deep learning has emerged as a transformative approach in medical image analysis, offering promising capabilities for enhancing cancer detection and diagnosis. Despite its potential, deep learning models often suffer from overfitting, which can substantially degrade their performance, especially when data variability is limited. To address these challenges, this study proposes a novel and comprehensive framework for lung cancer classification using computed tomography (CT) images, which involves several key stages. The framework begins with a data augmentation, applying techniques such as rotation, flipping, and scaling to improve data diversity and model robustness. Preprocessing is then carried out using median filtering to suppress image noise and enhance quality for subsequent analysis. Subsequently, an Improved SegNet (ISN) model is introduced for segmentation, incorporating a Mixed Average Lp (MALp) pooling layer. This proposed pooling mechanism enhances the retention of spatial information and improves the segmentation accuracy and efficiency. Then, pertinent features such as Modified Pyramid Histogram of Oriented Gradients (MPHOG), color and shape descriptors such as epsilon, area, perimeter, and hull are obtained from the segmented outcome, along with deep features extracted using the ResNet, VGG-16, and GoogLeNet. The obtained feature set is subjected to the classification phase, where a hybrid deep learning model is proposed for classifying lung cancer that integrates an Improved LeNet with LinkNet models and produces the final classified outcomes as Normal, Benign, and Malignant. Moreover, the experimental results demonstrate the superiority of the proposed framework over traditional methods, achieving notable performance metrics including an accuracy of 0.981, precision of 0.975, and F-measure of 0.942. These results underscore the practical significance of the framework in supporting early and accurate lung cancer diagnosis, ultimately contributing to improved clinical outcomes and potentially saving the lives of patients.