A convolutional neural network-based system to classify patients using FDG PET/CT examinations.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Nuclear Medicine

Background:

• Growing use of FDG PET/CT in oncology necessitates automated detection systems.
• Increasing demand for AI to prevent human oversight and misdiagnosis in PET/CT interpretation.
• Need for reliable AI tools to classify FDG PET/CT findings.

Purpose of the Study:

• To develop a convolutional neural network (CNN)-based system for classifying whole-body FDG PET scans.
• To categorize FDG PET scans into three classes: benign, malignant, or equivocal.
• To evaluate the performance of the CNN system in both patient-based and region-based analyses.

Main Methods:

• Retrospective study of 3485 patients undergoing whole-body FDG PET/CT.
• Classification of cases into benign, malignant, or equivocal by a nuclear medicine physician.
• Development of a residual network (ResNet)-based CNN architecture for classification and region-specific analysis.

Main Results:

• CNN achieved high accuracy: 99.4% for benign, 99.4% for malignant, and 87.5% for equivocal classifications (patient-based).
• Region-based analysis showed high accuracy: 97.3% (head-and-neck), 96.6% (chest), 92.8% (abdomen), and 99.6% (pelvic region).
• The system demonstrated reliable classification across different anatomical areas.

Conclusions:

• The developed CNN-based system reliably classifies FDG PET images into benign, malignant, and equivocal categories.
• This AI system can serve as a valuable double-checking tool for physicians.
• The system has the potential to reduce oversight and prevent misdiagnosis in oncology imaging.