Scholarlink

<p>This study presents a comprehensive comparative analysis of 10 state-of-the-art lightweight deep learning architectures for lung histopathological image classification, encompassing Convolutional Neural Networks, Vision Transformers, and hybrid CNN–Transformer models. All architectures were evaluated under identical and fully deterministic experimental conditions on a balanced dataset of 15,000 images representing Lung Adenocarcinoma, Lung Squamous Cell Carcinoma, and benign lung tissue. Two different experimental settings were used to assess the performance of each model: training without data augmentation and training with augmentation techniques. Performance was assessed using Accuracy, F1-score, AUC, Sensitivity, Specificity, final training loss, average epoch time, inference time, and memory consumption. Among all evaluated models, the CNN-based GhostNet-100 achieved the highest overall performance in both experimental settings. Without data augmentation, it reached 99.95% accuracy and a 99.95% F1-score, while with data augmentation it achieved 99.96% accuracy and a 99.96% F1-score. Hybrid architectures such as LeViT-128S also demonstrated competitive performance, whereas lightweight Vision Transformers, including ViT-Tiny and EVA-Tiny, exhibited comparatively lower performance across both settings. Overall, the findings demonstrate that GhostNet-100 provides the optimal balance between predictive accuracy and computational efficiency under the controlled experimental conditions of this study, suggesting its potential as a candidate architecture for future clinical decision-support applications pending prospective clinical validation.</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

Comparative analysis of lightweight CNN, vision transformer and hybrid architectures for lung histopathological image classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Doğu Manalı,
  • Hasan Demirel

摘要

This study presents a comprehensive comparative analysis of 10 state-of-the-art lightweight deep learning architectures for lung histopathological image classification, encompassing Convolutional Neural Networks, Vision Transformers, and hybrid CNN–Transformer models. All architectures were evaluated under identical and fully deterministic experimental conditions on a balanced dataset of 15,000 images representing Lung Adenocarcinoma, Lung Squamous Cell Carcinoma, and benign lung tissue. Two different experimental settings were used to assess the performance of each model: training without data augmentation and training with augmentation techniques. Performance was assessed using Accuracy, F1-score, AUC, Sensitivity, Specificity, final training loss, average epoch time, inference time, and memory consumption. Among all evaluated models, the CNN-based GhostNet-100 achieved the highest overall performance in both experimental settings. Without data augmentation, it reached 99.95% accuracy and a 99.95% F1-score, while with data augmentation it achieved 99.96% accuracy and a 99.96% F1-score. Hybrid architectures such as LeViT-128S also demonstrated competitive performance, whereas lightweight Vision Transformers, including ViT-Tiny and EVA-Tiny, exhibited comparatively lower performance across both settings. Overall, the findings demonstrate that GhostNet-100 provides the optimal balance between predictive accuracy and computational efficiency under the controlled experimental conditions of this study, suggesting its potential as a candidate architecture for future clinical decision-support applications pending prospective clinical validation.