A diagnostic classification of lung nodules using multiple-scale residual network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Computed tomography (CT) scans are crucial for lung cancer screening and mortality reduction.
• Accurate differentiation between benign and malignant pulmonary nodules on CT images remains a significant clinical challenge.

Purpose of the Study:

• To develop and evaluate a multiple-scale residual network (MResNet) for automated lung nodule classification using deep learning.
• To enhance the precision of extracting general features from lung nodules in CT images.

Main Methods:

• The MResNet model integrates residual units and a pyramid pooling module (PPM) for feature learning.
• ResNet serves as the backbone for contextual information and feature representation, while PPM fuses multi-scale features.

Main Results:

• The MResNet achieved high performance on the training set with 99.12% accuracy, 98.64% sensitivity, and 0.9998 AUC.
• On the testing set, MResNet demonstrated 85.23% accuracy, 92.79% sensitivity, and 0.9275 AUC, indicating strong performance in malignancy risk estimation.

Conclusions:

• The developed MResNet shows significant potential for accurately assessing the malignancy risk of pulmonary nodules detected via CT.
• This deep learning model can offer reliable malignancy risk scores, potentially optimizing lung cancer screening and management for clinicians.