Development of Combination Methods for Detecting Malignant Uptakes Based on Physiological Uptake Detection Using...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Deep learning is increasingly utilized in medical imaging analysis.
• Accurate detection of physiological and abnormal uptake in FDG-PET scans is crucial for diagnosis.
• Existing methods may require significant radiologist expertise and time.

Purpose of the Study:

• To apply the YOLOv2 object detection model for identifying physiological uptake in FDG-PET images.
• To develop and evaluate a combined technique with YOLOv2 for detecting abnormal uptake.
• To assess the precision and efficiency of YOLOv2 and the combined method as a diagnostic tool.

Main Methods:

• Trained YOLOv2 using transfer learning on 3,500 maximum intensity projection (MIP) FDG-PET images from 500 cases.
• Created a dataset by manually annotating physiological uptake regions of interest.
• Developed a combined method by subtracting YOLOv2-detected physiological uptake to highlight abnormalities and evaluated performance using coverage, false-positive, and false-negative rates.

Main Results:

• YOLOv2 achieved high average precision (AP) for physiological uptakes: brain (0.993), liver (0.913), and bladder (0.879).
• Mean average precision (mAP) was 0.831 across all classes with an IoU threshold of 0.5; average FPS was 31.60.
• The combined method demonstrated a high coverage rate (0.9205 ± 0.0312) and acceptable false-negative rate (0.1000 ± 0.0774) for abnormal uptake detection.

Conclusions:

• YOLOv2 effectively and rapidly detects physiological FDG-PET uptake on MIP images.
• The combined method leverages YOLOv2's capabilities to detect radiologist-identified abnormalities with high coverage.
• This approach shows promise as an efficient and precise diagnostic tool in nuclear medicine.