A 3D image segmentation for lung cancer using V.Net architecture based deep convolutional networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate lung segmentation in CT scans is vital for lung cancer identification and other diagnostic procedures.
• Distinguishing tumors from lung parenchyma requires precise segmentation algorithms.
• Fully convolutional networks (FCNs) offer a promising approach for semantic lung segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for accurate 3D lung segmentation in CT images.
• To improve the segmentation of tumors and lung parenchyma for enhanced cancer diagnosis.
• To assess the efficacy of a V-Net inspired FCN for this task.

Main Methods:

• Utilized CT cancer scans from the Task06_Lung database (64 training, 32 testing images).
• Applied a fully convolutional network (FCN) inspired by V-Net architecture for 3D segmentation.
• Implemented a system involving data preprocessing, data augmentation, and the V-Net model.

Main Results:

• Achieved an average Dice Score Coefficient (DSC) of 80% for region of interest (tumor) segmentation.
• Obtained an average DSC of 98% for surrounding lung tissue segmentation.
• Demonstrated superior performance compared to previous 3D lung segmentation methods.

Conclusions:

• The proposed 3D segmentation system provides robust and accurate lung segmentation, essential for tumor estimation.
• The V-Net based approach effectively segments both lung parenchyma and tumor regions.
• This method enhances the precision of lung cancer diagnosis through improved CT image analysis.