A generative whole-brain segmentation model for positron emission tomography images

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Positron emission tomography (PET) imaging is vital for neuroscience and clinical applications, offering insights into brain metabolism and activity.
• Low resolution of PET images can limit the accuracy of brain structure segmentation.
• Accurate segmentation is essential for understanding regional brain function and disease states.

Purpose of the Study:

• To develop an automatic and accurate whole-brain segmentation method for PET images.
• To address the limitations of low PET image resolution in segmenting brain structures.
• To propose a generative multi-object segmentation model tailored for brain PET imaging.

Main Methods:

• A two-protocol generative multi-object segmentation model was developed.
• A latent mapping model was pretrained to learn PET-MR image relationships, extracting anatomical information.
• A 3D segmentation model was applied to MR images, incorporating a custom cross-attention module for fusing functional and structural data.

Main Results:

• The proposed method demonstrated superior performance over existing deep learning approaches.
• Quantitative metrics included Dice (75.53% ± 4.26%), Jaccard (66.02% ± 4.55%), recall (74.64% ± 4.15%), and precision (81.40% ± 2.30%).
• The model accurately distinguished metabolic regions and showed clinical applicability.

Conclusions:

• A novel 3D generative multi-object segmentation model achieves accurate whole-brain segmentation for PET images.
• The proposed method outperforms other deep learning techniques in segmentation accuracy.
• Future work includes clinical application and extension to multimodal tasks.