Image-to-Biomarker Prediction of CSF Amyloid-β via Multimodal Deep Learning on MRI and Brain Perfusion SPECT

Alzheimer’s disease (AD) is the most common form of dementia. Recently, anti-amyloid-β (Aβ) antibody therapy has been applied to treat early AD. Therefore, the diagnosis of early AD has become more important than ever before. Cerebrospinal fluid (CSF) biomarkers such as amyloid-β (Aβ) and tau proteins are widely recognized indicators of AD pathology. Lumbar puncture to obtain CSF is invasive, and this procedure tends to be unfavorable for patients and their families. Although amyloid positron emission tomography (PET) is another method to confirm AD pathology, it is expensive. Single photon emission computed tomography (SPECT), which shows regional cerebral blood flow, has the advantage of being relatively inexpensive and can be used for repeated examination under medical insurance coverage in dementia clinical practice. To address this limitation, we present an image-to-biomarker deep learning framework for non-invasive prediction of CSF Aβ levels by leveraging multimodal neuroimaging. Our method integrates structural information from T1-weighted MRI and functional information from brain perfusion SPECT. All images are spatially normalized using DARTEL and processed using a ResNet-based 3D convolutional neural network (CNN). A multitask regression model is employed to simultaneously predict Aβ42, Aβ40, and the Aβ42/40 ratio, which is clinically valued for its robustness to individual variability, with age incorporated as an auxiliary input to improve prediction performance. The model is trained using a weighted mean squared error loss function and evaluated through five-fold cross-validation on a dataset of 153 patients with dementia. Ablation studies demonstrate that including both brain perfusion SPECT and age information significantly improves prediction performance, achieving root mean squared errors (RMSEs) of 500 pg/mL for Aβ42, 3,838 pg/mL for Aβ40, and 0.0355 for the Aβ42/Aβ40 ratio. These findings underscore the potential of our image-to-biomarker approach as a practical and accurate solution for CSF biomarker prediction, offering promising applications in improving the accuracy of early AD diagnosis, clinical decision-making support, and reducing the psychological, physical, and financial burden on patients and their families.