¹⁸F-FDG PET-based liver segmentation using deep-learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Liver segmentation traditionally relies on CT/MRI, facing alignment and artifact issues.
• Deep learning (DL) segmentation using 18F-FDG PET alone is underexplored.
• Accurate liver segmentation is crucial for assessing metabolic activity and treatment response.

Purpose of the Study:

• To develop and validate a DL model for segmenting the entire liver exclusively from 18F-FDG PET images.
• To overcome limitations of multi-modal imaging in liver segmentation.
• To enable rapid and stable evaluation of liver uptake from PET data.

Main Methods:

• Utilized a 3D U-Net architecture from nnUNet for segmentation.
• Trained and validated the model on 120 patient 18F-FDG PET datasets using 5-fold cross-validation.
• Evaluated segmentation accuracy with Intersection over Union (IoU) and Dice coefficient, and image quality with SUVmean, SUVmax, and SNR.

Main Results:

• Achieved high segmentation accuracy with an average IoU of 0.89 and Dice coefficient of 0.94 on the test set.
• Demonstrated no significant discrepancies in image quality metrics compared to ground truth.
• Successfully extracted liver regions from 18F-FDG PET images, enabling accurate uptake assessment.

Conclusions:

• A DL model can accurately segment the liver using only 18F-FDG PET images.
• This method offers a viable alternative to CT/MRI-based segmentation, reducing artifacts and alignment issues.
• The approach facilitates efficient and reliable evaluation of liver metabolic activity in clinical settings.