The application of deep learning in lung cancerous lesion detection

Abstract Background Characterized by rapid metastasis and a significant death rate, lung cancer presents a formidable challenge, which underscores the critical role of early detection in combating the disease. This study addresses the urgent need for early lung cancer detection using deep learning models applied to computed tomography (CT) images. Methods Our study introduced a unique non-cancer pneumonia dataset, a publicly available large-scale collection of high-quality pneumonia CT scans with detailed descriptions. We utilized this dataset to fine-tune nine pretrained models, including DenseNet121, MobileNetV2, InceptionV3, InceptionResNetV2, ResNet50, ResNet101, VGG16, VGG19, and Xception for the classification of lung cancer and pneumonia. Results ResNet50 demonstrated the highest accuracy and sensitivity (97.7% and 100%, respectively), while InceptionV3 excelled in precision (97.9%) and specificity (98.0%). The study also highlighted the contribution of the gradient-weighted class activation mapping (Grad-CAM) technique in examining the effectiveness of the model-training process via the visualization of features learned across different layers. Grad-CAM revealed that among the best-performed models, InceptionV3 successfully identified cancerous lesions in CT scans. Our findings demonstrated the potential of deep learning models in early lung cancer screening and improving the accuracy of the diagnosis procedure. Data availability The pneumonia CT scan dataset used in this study is extracted from peer-reviewed publications and can be accessed at https://github.com/ReiCHU31/CT-pneumonia-dataset