Anomaly detection in chest ¹⁸F-FDG PET/CT by Bayesian deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Positron Emission Tomography/Computed Tomography (PET/CT) is crucial for diagnosing various conditions.
• Detecting abnormal radiotracer uptake, such as with 18F-fluorodeoxyglucose (FDG), is key to identifying disease.
• Developing automated systems for anomaly detection can improve diagnostic efficiency and accuracy.

Purpose of the Study:

• To create an anomaly detection system for PET/CT using 18F-FDG.
• The system should be trained exclusively on normal PET/CT images.
• It must be capable of detecting abnormal FDG uptake anywhere within the chest region.

Main Methods:

• A Bayesian deep learning framework was employed for model training.
• The model was trained on a large dataset of 1878 normal PET/CT scans.
• Evaluation involved 34 scans with known focal abnormal FDG uptake in the chest (28 pulmonary, 17 extrapulmonary).

Main Results:

• The model achieved an area under the ROC curve of 0.992 for abnormal voxels and 0.852 for abnormal slices.
• It detected 91.1% of abnormal FDG uptake foci (41 out of 45).
• Sensitivity was 82.2% at 3.0 false positives per scan, identifying most pulmonary and extrapulmonary abnormalities.

Conclusions:

• A deep learning model trained solely on normal PET/CT scans can effectively detect abnormal FDG uptake.
• The system successfully identified both pulmonary and extrapulmonary abnormalities in the chest region.
• This approach offers a promising method for anomaly detection in PET/CT imaging without requiring abnormal training data.