Volumetric lung nodule segmentation using adaptive ROI with multi-view residual learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate pulmonary nodule quantification is crucial for early lung cancer diagnosis and patient survival.
• Existing segmentation methods using fixed regions of interest (ROIs) or volumes of interest (VOIs) have limitations in accuracy and scope.
• These limitations include restricted nodule investigation and inclusion of non-nodular structures, hindering precise segmentation.

Purpose of the Study:

• To propose a novel semi-automated approach for accurate 3D segmentation of lung nodules in CT scans.
• To overcome the limitations of traditional methods by enabling dynamic ROI selection and comprehensive nodule exploration.
• To enhance the accuracy of pulmonary nodule segmentation for improved lung cancer diagnosis.

Main Methods:

• A two-stage semi-automated segmentation technique was developed.
• Stage 1: Adaptive ROI algorithm and Deep Residual U-Net for patch-wise exploration along the axial axis, generating an initial nodule estimation and VOI.
• Stage 2: Residual U-Nets for patch-wise exploration along coronal and sagittal axes within the extracted VOI, followed by a consensus module for final volumetric segmentation.

Main Results:

• The proposed algorithm achieved an average Dice score of 87.5% on the LIDC-IDRI dataset.
• This performance is significantly higher than existing state-of-the-art segmentation techniques.
• The method demonstrated robust performance in accurately segmenting pulmonary nodules.

Conclusions:

• The novel semi-automated approach significantly improves the accuracy of 3D pulmonary nodule segmentation.
• This technique enhances the potential for early lung cancer diagnosis through precise nodule quantification.
• The developed method represents a substantial advancement over current state-of-the-art segmentation techniques.