Scholarlink

<p>Lung cancer continues to be a major cause of cancer-related death globally, highlighting the necessity for precise and prompt detection techniques. This paper presents FusionNet-ViT, a novel hybrid deep learning framework that merges EfficientNetB0 and Vision Transformers (ViT) for robust lung cancer classification from histopathology images. The proposed pipeline addresses reproducibility by explicitly detailing hyperparameters, data preprocessing, and augmentation strategies. EfficientNetB0 exploits convolutional layers to capture fine-grained structural features, while the ViT branch learns long-range dependencies through self-attention. Feature maps from both streams are concatenated in a fusion layer before final classification through a multi-layer perceptron. We evaluate our method using the publicly available LC25000 dataset, focusing on three subtypes of lung tissue adenocarcinoma, squamous cell carcinoma, and benign across 15,000 images. Experimental results show a 99.36% accuracy and near-perfect AUC (1.00), surpassing several state-of-the-art techniques. Detailed confusion matrix analyses and improvement rates relative to baselines further affirm the model’s strong discriminative capability. The results indicate that combining localized feature extraction with global context modeling leads to highly reliable lung cancer detection.</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

FusionNet-ViT: Hybrid Deep Learning Model for Lung Cancer Classification

  • 期刊论文

  • 发表日期：

    2025年11月5日

查看全文
  • Onkar Singh

摘要

Lung cancer continues to be a major cause of cancer-related death globally, highlighting the necessity for precise and prompt detection techniques. This paper presents FusionNet-ViT, a novel hybrid deep learning framework that merges EfficientNetB0 and Vision Transformers (ViT) for robust lung cancer classification from histopathology images. The proposed pipeline addresses reproducibility by explicitly detailing hyperparameters, data preprocessing, and augmentation strategies. EfficientNetB0 exploits convolutional layers to capture fine-grained structural features, while the ViT branch learns long-range dependencies through self-attention. Feature maps from both streams are concatenated in a fusion layer before final classification through a multi-layer perceptron. We evaluate our method using the publicly available LC25000 dataset, focusing on three subtypes of lung tissue adenocarcinoma, squamous cell carcinoma, and benign across 15,000 images. Experimental results show a 99.36% accuracy and near-perfect AUC (1.00), surpassing several state-of-the-art techniques. Detailed confusion matrix analyses and improvement rates relative to baselines further affirm the model’s strong discriminative capability. The results indicate that combining localized feature extraction with global context modeling leads to highly reliable lung cancer detection.