A Two-Stage Convolutional Neural Networks for Lung Nodule Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early lung cancer detection significantly improves patient survival rates.
• Accurate lung nodule detection in computed tomography (CT) images is crucial for lung cancer diagnosis.
• Challenges in lung nodule detection arise from nodule heterogeneity and complex surrounding environments.

Purpose of the Study:

• To propose a robust two-stage convolutional neural network (TSCNN) for accurate lung nodule detection.
• To enhance the recall rate while minimizing false positives in initial nodule detection.
• To reduce false positives effectively in the subsequent classification stage.

Main Methods:

• A two-stage approach utilizing an improved U-Net segmentation network for initial nodule detection.
• Implementation of a novel sampling strategy and a two-phase prediction method in the first stage.
• A second stage employing a dual pooling structure with three 3D-CNNs for false positive reduction.
• Data augmentation using random masks and ensemble learning to improve model generalization.

Main Results:

• The proposed TSCNN architecture demonstrated competitive detection performance on the LUNA dataset.
• The two-stage approach effectively balanced high recall with reduced false positives.
• Ensemble learning enhanced the generalization capability of the false positive reduction model.

Conclusions:

• The TSCNN architecture offers a promising solution for robust lung nodule detection.
• The proposed methods address key challenges in lung nodule identification from CT images.
• This approach contributes to advancing early lung cancer diagnosis through improved imaging analysis.