Scholarlink

<p>Lung cancer accounts for the highest number of deaths due to cancer in the world. Precise and prompt identification on the basis of histopathology pictures is vital in enhancing survival outcomes. This study proposes to create a fast and intelligible CNN with high accuracy and an easy-to-use CBQA model that can be deployed in the clinical setting. To do this, we created a bespoke model of CNN with four convolutional layers, group normalisation, dropout and SoftMax output. A small balanced dataset of 15,000 histopathology images (5000 adenocarcinoma, 5000 squamous cell carcinoma, 5000 benign) out of Kaggle dataset of Lung Cancer Histopathological Images was adopted with a 70–15–15% train–validation–test split. Data augmentation was used, and TCGA-LUAD patches were validated externally. Benchmarking comprised the classical machine learning models (SVM, Logistic Regression, Random Forest) and popular CNNs (EfficientNetB3, XLLC-Net). The proposed CNN demonstrated 95.3% accuracy, 94.2% precision, 93.5% recall, 93.8% F1, and 0.965 AUC-ROC on LC25000, yielding a superior performance as compared to SVM (86.7%), Logistic Regression (84.3%) and Random Forest (85.2%). Performance can be seen to be competitive with EfficientNetB3 (94.1%) and XLLC-Net (95.0%), with a far lesser use of computational resources. External validation on TCGA-LUAD gave an accuracy of 91.2%. Nucleus-dense regions were also observed on, validating it in clinical practice. In contrast to previous studies, which have to a large extent employed massive pretrained networks, this study proposes a lightweight, easy-to-interpret, and externally curated CNN that is shown to perform in real-time (38&#xa0;ms per picture) and performs well in generalisation. The synergy that allows using balanced dataset training, external validation, and interpretability allows assuring the clinical applicability.</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

Revolutionising Early Diagnosis of Lung Cancer: A Deep Learning Framework for Precision Detection

  • 期刊论文

  • 发表日期：

    2026年5月17日

查看全文
  • Kaushik Roy Choudhury,
  • Sanket Dan,
  • Bikramjit Sarkar,
  • Debarpan Das,
  • Dipeswar Pal

摘要

Lung cancer accounts for the highest number of deaths due to cancer in the world. Precise and prompt identification on the basis of histopathology pictures is vital in enhancing survival outcomes. This study proposes to create a fast and intelligible CNN with high accuracy and an easy-to-use CBQA model that can be deployed in the clinical setting. To do this, we created a bespoke model of CNN with four convolutional layers, group normalisation, dropout and SoftMax output. A small balanced dataset of 15,000 histopathology images (5000 adenocarcinoma, 5000 squamous cell carcinoma, 5000 benign) out of Kaggle dataset of Lung Cancer Histopathological Images was adopted with a 70–15–15% train–validation–test split. Data augmentation was used, and TCGA-LUAD patches were validated externally. Benchmarking comprised the classical machine learning models (SVM, Logistic Regression, Random Forest) and popular CNNs (EfficientNetB3, XLLC-Net). The proposed CNN demonstrated 95.3% accuracy, 94.2% precision, 93.5% recall, 93.8% F1, and 0.965 AUC-ROC on LC25000, yielding a superior performance as compared to SVM (86.7%), Logistic Regression (84.3%) and Random Forest (85.2%). Performance can be seen to be competitive with EfficientNetB3 (94.1%) and XLLC-Net (95.0%), with a far lesser use of computational resources. External validation on TCGA-LUAD gave an accuracy of 91.2%. Nucleus-dense regions were also observed on, validating it in clinical practice. In contrast to previous studies, which have to a large extent employed massive pretrained networks, this study proposes a lightweight, easy-to-interpret, and externally curated CNN that is shown to perform in real-time (38 ms per picture) and performs well in generalisation. The synergy that allows using balanced dataset training, external validation, and interpretability allows assuring the clinical applicability.