Automatic Extraction of PET RANO Criteria with an Externally Validated Deep Learning Model: Application to [18F]FDOPA...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate glioma segmentation on amino acid PET is crucial for quantitative tumor assessment during treatment.
• This study focuses on developing an automated deep learning model for extracting PET RANO criteria from [18F]FDOPA PET scans.

Purpose of the Study:

• To develop and validate a deep learning model for automated segmentation of gliomas on [18F]FDOPA PET.
• To assess the model's performance in extracting quantitative PET RANO criteria.

Main Methods:

• A 3D U-Net model was trained on 530 [18F]FDOPA PET scans from three European centers.
• The model segmented tumor and healthy brain volumes, with performance evaluated using Dice coefficients.
• Quantitative agreement for PET RANO criteria, TBRs, and MTV was assessed at the lesion level.

Main Results:

• The model achieved high Dice coefficients (0.851–0.925) across training, validation, and test sets.
• Excellent agreement was found with expert quantification for MTV, TBRmax, and TBRmean (ICC > 0.93).
• Over 97% of measurable lesions were correctly identified.

Conclusions:

• The developed [18F]FDOPA PET deep learning model shows robust multicenter performance.
• The model enables fully automated and reproducible quantification of gliomas.
• This supports broader clinical adoption of amino-acid PET in neuro-oncology.