Deep learning signature of brain [18F]FDG PET associated with cognitive outcome of rapid eye movement sleep behavior disorder

Abstract

An objective biomarker to predict the outcome of isolated rapid eye movement sleep behavior disorder (iRBD) is crucial for the management. This study aimed to investigate cognitive signature of brain [18F]FDG PET based on deep learning (DL) for evaluating patients with iRBD. Fifty iRBD patients, 19 with mild cognitive impairment (MCI) (RBD-MCI) and 31 without MCI (RBD-nonMCI), were prospectively enrolled. A DL model for the cognitive signature was trained by using Alzheimer’s Disease Neuroimaging Initiative database and transferred to baseline [18F]FDG PET from the iRBD cohort. The results showed that the DL-based cognitive dysfunction score was significantly higher in RBD-MCI than in RBD-nonMCI. The AUC of ROC curve for differentiating RBD-MCI from RBD-nonMCI was 0.70 (95% CI 0.56–0.82). The baseline DL-based cognitive dysfunction score was significantly higher in iRBD patients who showed a decrease in CERAD scores during 2 years than in those who did not. Brain metabolic features related to cognitive dysfunction-related regions of individual iRBD patients mainly included posterior cortical regions. This work demonstrates that the cognitive signature based on DL could be used to objectively evaluate cognitive function in iRBD. We suggest that this approach could be extended to an objective biomarker predicting cognitive decline and neurodegeneration in iRBD.

Deep learning-based multimodality classification of chronic mild traumatic brain injury using resting-state functional MRI and PET imaging

“…Similar to our study, Zou et al incorporated several rs-fMRI metrics including fALFF, ReHo, and VMHC into a 3D CNN model to differentiate subjects with attention deficit hyperactivity disorder (ADHD) from normal subjects and reported a mean classification accuracy of 66.04 and 69.15% for single modality and multimodality 3D CNN, respectively (Zou et al, 2017). Similarly, Ghanbari et al employed single and combined rs-fMRI metrics including fALFF, ReHo, and VMHC in a classification model of 3D CNN to identify patients with schizophrenia from HCs and found the accuracy of 72.20, 79.55, 87.63, and 90.91% for fALFF, ReHo, VMHC, and multimodality models, respectively (Ryoo et al, 2022). While many studies have been carried out on the classification between the patient and control groups using conventional ML algorithms, some recent studies have shown the capability and superiority of DNN with several hidden layers to extract lower-to-higher level information through several hidden layers.…”