Accurate segmentation for different types of lung nodules on CT images using improved U-Net convolutional network

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Accurate lung nodule segmentation in computed tomography (CT) images is challenging due to variations in nodule characteristics and attachments.
• Existing segmentation methods struggle with the diverse appearances and complex anatomical contexts of lung nodules.

Purpose of the Study:

• To develop an accurate lung nodule segmentation method using an improved U-Net convolutional network.
• To enhance the segmentation of various lung nodule types in CT images.

Main Methods:

• A two-phase approach involving lung parenchyma segmentation with the α-hull algorithm and nodule segmentation using an improved U-Net with batch normalization.
• Utilized Dice loss for superior performance compared to mean square error and Binary_crossentropy loss.

Main Results:

• The improved U-Net model achieved a Dice similarity coefficient of 0.8623, outperforming existing state-of-the-art algorithms.
• The α-hull algorithm and batch normalization significantly improved segmentation accuracy.
• Dice loss demonstrated superior segmentation performance.

Conclusions:

• The proposed improved U-Net network effectively segments diverse lung nodules in CT images.
• This method offers practical value for radiologists in lung nodule segmentation and lung cancer diagnosis.
• The integration of α-hull and batch normalization enhances segmentation efficacy.