Semantic Segmentation of White Matter in FDG-PET Using Generative Adversarial Network

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• ¹⁸F-FDG PET/CT detects early functional changes in neurodegenerative disorders.
• Low spatial resolution of PET/CT necessitates integration with anatomical imaging (CT/MRI) for interpretation.
• Lack of structural information hinders accurate segmentation and quantification of PET/CT scans.

Purpose of the Study:

• To develop a method for segmenting brain white matter directly from ¹⁸F-FDG PET/CT images.
• To enable quantitative analysis of ¹⁸F-FDG PET/CT scans without relying on separate anatomical modalities.
• To improve the diagnostic utility of ¹⁸F-FDG PET/CT in neurodegenerative disease assessment.

Main Methods:

• Proposed a generative adversarial network (GAN) for white matter segmentation in ¹⁸F-FDG PET/CT images.
• Evaluated segmentation performance using Dice, AUC-PR, precision, and recall metrics.
• Compared the proposed GAN method against other deep learning segmentation techniques.

Main Results:

• The proposed GAN model achieved superior segmentation accuracy compared to existing deep learning methods.
• Demonstrated high reliability in segmenting white matter from ¹⁸F-FDG PET/CT data.
• The method enables accurate quantitative analysis solely from PET/CT scans.

Conclusions:

• The developed GAN-based method effectively segments white matter in ¹⁸F-FDG PET/CT images.
• This approach overcomes the limitations of low spatial resolution and reliance on multimodal imaging.
• The findings support the use of this method for enhanced quantitative analysis in neurodegenerative disorder diagnosis.