A Deep Learning Model Predicts a Diagnosis of Alzheimer’s Disease Using 18F-FDG PET
Alzheimer’s disease remains a diagnosis made on clinical grounds. However, advancements in diagnostic technology such as positron emission tomography (PET) with 18F-fluorodeoxyglucose (18F-FDG) have created opportunities for earlier diagnosis and thus, more effective treatment. At present, 18F-FDG PET requires interpretation by specialists in nuclear medicine and neuroimaging based on pattern recognition using primarily qualitative reads. Within the radiology community, there is increasing recognition that deep learning may assist in addressing the increasing complexity and volume of such imaging data.
In a recent study, researchers from the UC San Francisco Department of Radiology and Biomedical Imaging aimed to develop a deep learning algorithm to predict the final clinical diagnoses of patients who underwent 18F-FDG PET of the brain. They then compared the performance of their deep learning algorithm with the current standard clinical reading methods. “Deep learning is ideally suited for such a problem because it is a nonlinear function estimator particularly strong at capturing diffuse yet subtle processes in images, such as in PET scans of Alzheimer’s disease,” says Jae Ho Sohn, MD, MS, the study’s corresponding author. Their paper was recently published in Radiology.
Such research has important implications for patient care. A deep learning algorithm can be used to improve the accuracy of diagnosing Alzheimer’s disease from 18F-FDG PET scans of the brain. By predicting Alzheimer’s disease earlier in the disease course in conjunction with other biochemical and imaging tests, a deep learning algorithm provides an opportunity for early therapeutic intervention. Dr. Sohn and team hypothesized that the deep learning algorithm could detect features or patterns that are not evident in standard clinical review of images and thereby improve the final diagnostic classification of individuals.
In their study, researchers collected 18F-FDG PET brain images from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and a retrospective independent test dataset from UCSF. Final clinical diagnosis on follow-up was recorded. They developed a deep learning algorithm that achieved 82 percent specificity at 100 percent sensitivity at predicting Alzheimer’s disease, an average of 75.8 months prior to the final diagnosis.
“Hands-on mentorship from a hybrid physician-engineer like Dr. Sohn through the Big Data in Radiology (BDRAD) team was really helpful in the development of a deep learning algorithm tailored to the medical domain,” says Yiming Ding, the first author and an undergraduate student at UC Berkeley. “I'm very excited to be making direct impact on patient lives through my engineering work.”
Dr. Sohn, a radiology resident leader of the BDRAD team, was in turn mentored by faculty advisors and senior authors, Youngho Seo, PhD and Benjamin Franc, MD. Other authors from UCSF included Michael Kawczynsk, MS; Hari Trivedi, MD; Roy Harnish, MS; Nathaniel Jenkins, MS; Dmytro Lituiev, PhD; Timothy Copeland, MPP; Mariam Aboian, MD, PhD; Carina Mari Aparici, MD; Spencer Behr, MD; Robert Flavell, MD, PhD; Shih-Ying Huang, PhD; Kelly Zalocusky, PhD; Lorenzo Nardo, PhD; Randall Hawkins, MD, PhD; Miguel Hernandez Pampaloni, MD, PhD; and Dexter Hadley, MD, PhD.