Deep learning only by normal brain PET identify unheralded brain anomalies

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning models show high diagnostic accuracy but struggle with real-world clinical data variability.
• A gap exists between training datasets and diverse clinical imaging, limiting model generalizability.
• Unsupervised learning offers a potential solution to bridge this gap using only normal data.

Purpose of the Study:

• To develop an unsupervised deep learning model trained exclusively on normal brain PET data.
• To assess the model's ability to detect abnormalities across various neurological disorders in clinical routine data.
• To evaluate the model's utility in assisting expert visual interpretation of brain PET scans.

Main Methods:

• Utilized a variational autoencoder for unsupervised learning to define an "Abnormality Score."
• Applied the model to FDG PET data from Alzheimer's disease (AD), mild cognitive impairment (MCI), and clinical routine cases.
• Measured diagnostic accuracy using the area under the curve (AUC) of receiver-operating characteristic (ROC) curves.

Main Results:

• Achieved an AUC of 0.90 for differentiating AD and showed significant correlation between Abnormality Score and cognitive decline.
• Attained an AUC of 0.74 for discriminating various disorders from controls.
• Enhanced expert visual interpretation, identifying abnormal patterns in 60% of cases missed initially.

Conclusions:

• The unsupervised deep learning model trained on normal data effectively identifies diverse brain abnormalities, including rare conditions.
• This approach shows promise for interpreting real-world clinical brain PET data.
• The model aids in the detection of abnormalities, potentially improving diagnostic efficiency and accuracy.