Automatic localization of normal active organs in 3D PET scans.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Positron Emission Tomography (PET) imaging is crucial for cancer detection and treatment evaluation.
• Visual interpretation of PET scans requires distinguishing normal organ activity from abnormal tumor growth.
• Automating this distinction is essential for efficient and accurate clinical workflows.

Purpose of the Study:

• To develop a deep learning method for localizing and detecting normal active organs in 3D PET scans.
• To enable automated differentiation between physiological organ uptake and pathological tumor activity.
• To improve the reliability and efficiency of PET scan analysis in oncology.

Main Methods:

• Adaptation of the YOLO (You Only Look Once) deep network architecture for multi-organ detection in 2D PET slices.
• Aggregation of 2D detection results to generate semantically labeled 3D bounding boxes for organs.
• Evaluation on a dataset of 479 18F-FDG PET scans from 156 patients.

Main Results:

• Achieved high average organ detection precision (75-98%) and recall (94-100%).
• Demonstrated accurate localization with average bounding box centroid error < 14 mm and wall error < 24 mm.
• Obtained a mean Intersection over Union (IOU) of up to 72% for organ detection.

Conclusions:

• The proposed deep learning method effectively localizes and detects normal organs in 3D PET scans.
• This automated approach can assist clinicians in differentiating normal physiological activity from tumor uptake.
• The findings support the potential of AI in enhancing cancer diagnosis and treatment monitoring using PET imaging.