Unsupervised class labeling of diffuse lung diseases using frequent attribute patterns.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Mining

Background:

• Computer-aided diagnosis (CAD) for CT images requires accurate classifiers.
• Training classifiers necessitates large datasets with radiologist-assigned labels.
• Manual labeling of CT images is time-consuming and impractical for radiologists.

Purpose of the Study:

• To develop an unsupervised class labeling mechanism for diffuse lung diseases in CT images.
• To propose a novel clustering algorithm that does not rely on correct labels.
• To address the challenge of acquiring labeled data for CAD systems.

Main Methods:

• Extraction of frequent opacity patterns using genetic network programming (GNP).
• Automatic distribution of extracted patterns into clusters using a genetic algorithm (GA).
• Application to lung CT images for clustering normal and diffuse lung diseases.

Main Results:

• 1,148 frequent attribute patterns were extracted by GNP.
• GA was employed to create clusters for normal and five types of abnormal opacities (six-class problem).
• The unsupervised method achieved 47.7% clustering accuracy.

Conclusions:

• The proposed method successfully creates clusters without requiring correct labels.
• This approach shows potential for application in CAD systems.
• It can significantly reduce the time and cost associated with labeling CT images.