Reduction of false positives in lung nodule detection using a two-level neural classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computer-aided diagnosis (CAD) systems are crucial for detecting lung cancer nodules on digitized chest radiographs.
• Existing systems often struggle with high false-positive rates, impacting diagnostic efficiency.

Purpose of the Study:

• To present a two-level neural classification method for reducing false positives in lung cancer nodule detection.
• To evaluate the performance of a parameterized two-level Convolutional Neural Network (CNN) architecture for this task.

Main Methods:

• Development of a parameterized two-level CNN architecture with a special multilabel output encoding procedure.
• Training and evaluation on digitized chest radiographs for automatic nodule localization, feature extraction, and diagnosis.
• Utilizing the Receiver Operating Characteristic (ROC) method with the area under the curve (A(z)) as the performance index.

Main Results:

• The two-level CNN achieved a superior performance index of A(z)=0.93.
• This significantly outperformed a single-level CNN, which achieved A(z)=0.85.
• The system demonstrated high true-positive fraction detection and low false-positive fraction detection.

Conclusions:

• The proposed two-level CNN architecture is a promising approach for reducing false positives in lung cancer diagnosis.
• The architecture is extensible, problem-independent, and applicable to other 2-D medical imaging diagnostic tasks.