Deep learning‐based cortical thickness maps for diagnosis of neurodegenerative diseases: a rater study

Abstract INTRODUCTION Cortical atrophy patterns on magnetic resonance imaging (MRI) are essential for diagnosing neurodegenerative diseases (NDs), but remain challenging to assess visually. Deep learning enables quantitative evaluation of cortical thickness (CTh) and z‐score maps.

Methods

3T three‐dimensional T1 MRI from 40 ND patients (Alzheimer's dementia, posterior cortical atrophy, behavioral variant frontotemporal dementia, semantic variant primary progressive aphasia) and 10 controls were retrospectively analyzed. Cortical surfaces were extracted with FreeSurfer and registered to fsaverage, and z‐score maps were generated using stochastic cortical self‐reconstruction (SCSR). Three neuroradiologists rated CTh or z‐score maps, each with or without 3D T1 for ND presence and differential diagnosis. Conventional 3D T1 served as baseline reading condition.

Results

Diagnostic accuracy (Acc) was quantitatively highest for z‐score maps with 3D T1 (Acc = 0.98) for the detection of ND, though it did not reach statistical significance. However, diagnostic confidence improved for z‐score maps versus baseline 3D T1 (adjusted p = 0.017 for Rater 2). Inter‐rater agreement improved from κ = 0.525 (baseline 3D T1) to κ = 0.792 (z‐score with 3D T1).

Discussion

SCSR‐generated z‐score maps show promise for diagnostic evaluation of ND.