Advancing Pulmonary Nodule Classification: A Novel Multi-Scale Fusion and Joint Upsampling Strategy using 3D Convolutional Neural Networks
Lung cancer, renowned for having the highest global incidence and mortality rates among all cancers, presents a promising avenue for improving survival rates through early detection and precise diagnosis. However, current diagnostic methods relying on the manual interpretation of CT images are susceptible to subjectivity and potential errors. To address this challenge, we introduce an innovative fully convolutional neural network that synergistically integrates multi-scale feature fusion and joint upsampling modules. Our model aims to enhance the precision of lung cancer diagnosis by effectively categorizing benign and malignant pulmonary nodules within CT images. Leveraging a comprehensive dataset comprising 1012 pulmonary nodule samples sourced from LIDC-IDRI, our evaluation reveals exceptional performance metrics. Notably, the model achieves an impressive area under the ROC curve of 97.35%, along with a high accuracy of 94.21%, sensitivity of 93.79%, and specificity of 94.91%.
Paper
Full text
Advancing Pulmonary Nodule Classification: A Novel Multi-Scale Fusion and Joint Upsampling Strategy using 3D Convolutional Neural Networks
Semantic Scholar · Medicine · 2023
Abstract
Lung cancer, renowned for having the highest global incidence and mortality rates among all cancers, presents a promising avenue for improving survival rates through early detection and precise diagnosis. However, current diagnostic methods relying on the manual interpretation of CT images are susceptible to subjectivity and potential errors. To address this challenge, we introduce an innovative fully convolutional neural network that synergistically integrates multi-scale feature fusion and joint upsampling modules. Our model aims to enhance the precision of lung cancer diagnosis by effectively categorizing benign and malignant pulmonary nodules within CT images. Leveraging a comprehensive dataset comprising 1012 pulmonary nodule samples sourced from LIDC-IDRI, our evaluation reveals exceptional performance metrics. Notably, the model achieves an impressive area under the ROC curve of 97.35%, along with a high accuracy of 94.21%, sensitivity of 93.79%, and specificity of 94.91%.