Scholarlink

Background <p>Accurate differentiation between lung adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC) is essential for guiding treatment decisions and predicting outcomes for non-small cell lung cancer (NSCLC). However, CT-based subtype discrimination remains challenging due to overlapping imaging features and inter-observer variability.</p> Methods <p>We developed an improved deep learning framework integrating residual connections, attention mechanisms, and a bidirectional feature pyramid network (BiFPN), combined with an ensemble strategy, to classify NSCLC subtypes on annotated CT slices. A total of 290 patients were retrospectively enrolled from two centers, with 200 used for training and internal validation, and 90 for external testing. Ablation experiments were conducted to evaluate the contribution of each architectural component. Model performance was assessed using F1-score, precision, accuracy, and the area under the receiver operating characteristic curve (AUC). Grad-CAM was applied to provide visual interpretability.</p> Results <p>In the internal validation cohort, the proposed model achieved an F1-score of 0.9973, precision of 0.9946, accuracy of 0.995, and an AUC of 1.0000, outperforming all benchmark models. In the external test cohort, the model maintained robust generalizability with an F1-score of 0.9508, precision of 0.9667, accuracy of 0.9333, and an AUC of 0.9891, surpassing competing methods. Grad-CAM visualizations demonstrated that the model predominantly focused on radiologically relevant regions, aligning with known imaging features of LUAD and LUSC.</p> Conclusions <p>This study presents a segmentation-guided, interpretable deep learning model for CT-based NSCLC subtype classification. The model may serve as an adjunctive tool for pre-treatment subtype assessment, while further validation using fully automated lesion localization and larger multicenter datasets is warranted before clinical implementation.</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

Multi-center validated attention-BiFPN deep learning for CT-based lung cancer subtype classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月3日

查看全文
  • Chutong Lin,
  • Haoran Jiang,
  • Shanwu Ma,
  • Jizheng Tang,
  • Yingze Ning,
  • Liang Jin,
  • Wei He,
  • Jie Bai,
  • Zhencheng Xiong,
  • Beichuan Zhu,
  • Guangliang Qiang

摘要

Background

Accurate differentiation between lung adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC) is essential for guiding treatment decisions and predicting outcomes for non-small cell lung cancer (NSCLC). However, CT-based subtype discrimination remains challenging due to overlapping imaging features and inter-observer variability.

Methods

We developed an improved deep learning framework integrating residual connections, attention mechanisms, and a bidirectional feature pyramid network (BiFPN), combined with an ensemble strategy, to classify NSCLC subtypes on annotated CT slices. A total of 290 patients were retrospectively enrolled from two centers, with 200 used for training and internal validation, and 90 for external testing. Ablation experiments were conducted to evaluate the contribution of each architectural component. Model performance was assessed using F1-score, precision, accuracy, and the area under the receiver operating characteristic curve (AUC). Grad-CAM was applied to provide visual interpretability.

Results

In the internal validation cohort, the proposed model achieved an F1-score of 0.9973, precision of 0.9946, accuracy of 0.995, and an AUC of 1.0000, outperforming all benchmark models. In the external test cohort, the model maintained robust generalizability with an F1-score of 0.9508, precision of 0.9667, accuracy of 0.9333, and an AUC of 0.9891, surpassing competing methods. Grad-CAM visualizations demonstrated that the model predominantly focused on radiologically relevant regions, aligning with known imaging features of LUAD and LUSC.

Conclusions

This study presents a segmentation-guided, interpretable deep learning model for CT-based NSCLC subtype classification. The model may serve as an adjunctive tool for pre-treatment subtype assessment, while further validation using fully automated lesion localization and larger multicenter datasets is warranted before clinical implementation.