Segmentation of lung parenchyma in CT images using CNN trained with the clustering algorithm generated dataset

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiology and Diagnostic Imaging

Background:

• Lung segmentation is crucial for early diagnosis and treatment of lung diseases.
• Accurate segmentation of lung parenchyma aids in locating and analyzing lung lesions.
• Machine learning approaches for lung parenchyma segmentation are underexplored.

Purpose of the Study:

• To develop an automated lung parenchyma segmentation method using machine learning.
• To create an efficient dataset generation pipeline for training segmentation models.
• To evaluate the performance of a proposed convolutional neural network (CNN) model for lung segmentation.

Main Methods:

• A clustering algorithm-based method was developed to generate training datasets for CNNs.
• K-means clustering was applied to CT image patches, followed by verification and expansion steps.
• A CNN architecture with a single convolutional layer, pooling, and fully connected layers was designed and trained.

Main Results:

• The CNN model achieved an average F-score of 0.9917 and an area under the curve of 0.9991.
• Accurate lung parenchyma segmentation was demonstrated across 201 subjects with diverse lung diseases and CT scanners.
• The proposed method achieved a high overlap ratio of 0.96 compared to manual segmentation.

Conclusions:

• The clustering algorithm effectively generates datasets for CNN-based lung segmentation.
• The developed CNN model provides highly accurate lung parenchyma segmentation.
• This method holds potential for improving the localization and analysis of lung lesions in clinical settings.