Automatic Lung Cancer Segmentation in [¹⁸F]FDG PET/CT Using a Two-Stage Deep Learning Approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiomics

Background:

• Accurate lung cancer segmentation is crucial for determining tumor functional volume in [18F]FDG PET/CT.
• Current segmentation methods may lack the precision required for detailed tumor analysis.

Purpose of the Study:

• To propose and evaluate a two-stage U-Net architecture for enhanced lung cancer segmentation in [18F]FDG PET/CT.
• To improve the accuracy and efficiency of tumor volume of interest (VOI) delineation.

Main Methods:

• Retrospective analysis of 887 whole-body [18F]FDG PET/CT scans.
• A two-stage U-Net model: Stage 1 (global U-Net) for preliminary tumor area extraction, Stage 2 (regional U-Net) for detailed segmentation using consecutive slices.
• Dataset split into training (730), validation (81), and testing (76) sets.

Main Results:

• The two-stage U-Net architecture demonstrated superior performance compared to a conventional one-stage 3D U-Net for primary lung cancer segmentation.
• The model accurately predicted detailed tumor margins.
• Quantitative analysis using Dice similarity coefficient confirmed the superiority of the two-stage approach.

Conclusions:

• The proposed two-stage U-Net method offers significant improvements in lung cancer segmentation accuracy.
• This approach is valuable for reducing the time and effort associated with precise tumor delineation in [18F]FDG PET/CT scans.
• Enhanced segmentation can lead to better functional volume assessment for lung cancer patients.