Automated deep learning segmentation of cardiac inflammatory FDG PET.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Fluorodeoxyglucose positron emission tomography (FDG PET) is crucial for diagnosing cardiac sarcoidosis.
• Manual image processing for FDG PET, including myocardial segmentation, is time-consuming.
• A 3D U-Net deep-learning (DL) algorithm was developed for automated myocardial segmentation.

Purpose of the Study:

• To develop and evaluate a DL algorithm for automated myocardial segmentation in cardiac sarcoidosis FDG PET.
• To compare the DL method with current automated segmentation techniques.
• To assess the impact of DL segmentation on processing time and clinical readability.

Main Methods:

• A 3D U-Net DL model was trained on FDG PET scans from 316 patients.
• Segmentation was validated against manual segmentation and current automated methods on a 50-patient subset.
• Clinical readability, left ventricle displacement and angulation, SUVmax correlation, and processing time were analyzed.

Main Results:

• DL segmentation improved readability scores in over 90% of cases compared to standard methods.
• DL segmentation performance was comparable to a trained technologist, outperforming standard segmentation.
• Using DL segmentation as initial placement significantly reduced manual segmentation processing time.

Conclusions:

• A novel DL-based automated segmentation tool significantly enhances cardiac sarcoidosis FDG PET processing.
• The tool provides optimized image display without user input.
• It offers substantial improvements in processing time for manual segmentation tasks.