Association of deep‐learning–derived brain computed tomography measures with cognition and blood‐based biomarkers of neurodegenerative diseases

Background Existing imaging biomarkers for neurodegenerative diseases such as magnetic resonance imaging (MRI) and positron emission tomography are costly or have limited accessibility. Computed tomography (CT) and novel blood tests have greater potential as screening modalities for suspected neurodegenerative diseases and to potentially exclude other treatable causes of neurodegeneration. Objective To evaluate the association of volumetric brain measures, derived using a novel CT‐based deep‐learning algorithm with cognition and blood levels of neurodegeneration markers. Method We previously developed deep‐learning‐based models using 734 datasets (722 cognitively normals (CN)) from Gothenburg H70 Birth Cohort to segment head CTs by learning from MRI‐derived labels. We used these models to segment, without further training, 204 novel CT datasets (73 Alzheimer's disease (AD), 49 mild cognitive impairment (MCI), 20 vascular dementia (VAD), 40 vascular cognitive impairment (VCI), 22 CN) recruited from the Singaporean Memory Clinic cohort (mean age= 74.0±8.2 years, 50.5% female), without MRI intervention. We tested associations between six CT‐derived volumetric measures (grey matter (GM), white matter (WM), GM/CSF, GM/ventricular CSF (VCSF), brain volume (BV)/CSF, BV/VCSF) with diagnosis, plasma biomarkers, and cognition. The plasma biomarkers were measured using a Simoa platform (Quanterix, Bilerica, MA,USA). Result The novel CT segmentations correlated significantly with their corresponding MRI‐segmentations. Lower CT‐derived tissue volumes correlated with higher plasma neurofilament (NFL) concentrations in the H70 cohort. In the MCC NUH cohort, lower CT‐derived volumes correlated with lower MMSE scores and higher plasma NFL. CT‐derived volumes yielded high diagnostic accuracy for AD vs CN (AUC ROC 0.87). Mean CT‐based volumetric measures were lower in AD and VAD compared with CN. Kruskal‐Wallis tests showed a statistically significant difference in CT‐based volumetric measures predominantly for GM/CSF and BV/CSF across all diagnostic groups, notably between AD vs CN and CN vs early stages of dementia. Conclusion Deep‐learning‐derived CT‐based volumetric measures correlate with relevant other imaging, cognitive and biochemical markers of neurodegenerative diseases. This supports the potential application of CT‐derived volumetric measures in aiding diagnostics and early detection of dementia. Further explorations of the association between CT‐based measures wi