3-D Lung Segmentation by Incremental Constrained Nonnegative Matrix Factorization.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Accurate lung segmentation in 3D chest computed tomography (CT) is vital for computer-assisted cancer diagnostics.
• Existing methods may struggle with lung tissue inhomogeneities and require extensive domain expertise.
• Scalability to large 3D datasets is a significant challenge.

Purpose of the Study:

• To develop an efficient and accurate 3D lung segmentation method for large CT images.
• To improve segmentation robustness in the presence of lung tissue inhomogeneities and pathologies.
• To reduce reliance on domain expert knowledge and simplify parameter tuning.

Main Methods:

• Utilized unsupervised learning to extract voxel-wise features of spatial image contexts.
• Proposed an incremental constrained nonnegative matrix factorization (ICNMF) approach.
• Incorporated smoothness constraints and slice-wise incremental learning with interslice dependency considerations.

Main Results:

• ICNMF demonstrated superior segmentation accuracy compared to state-of-the-art techniques, achieving a Dice similarity coefficient of 0.96 on in vivo data.
• The method showed robustness to lung tissue inhomogeneities and improved segmentation of internal lung pathologies.
• ICNMF achieved high rankings in the Lobe and Lung Analysis 2011 (LOLA11) study, with accuracy reaching 0.986 after excluding highly pathological cases.

Conclusions:

• The proposed ICNMF method offers an efficient, scalable, and accurate solution for 3D lung segmentation in CT images.
• ICNMF requires less domain expertise and fewer parameters, making it more adaptable.
• This technique holds significant potential for advancing computer-assisted lung cancer diagnostics.