Deep learning cortical z-score maps increase inter-rater agreement in dementia MRI reads

In a rater study of 40 neurodegeneration patients and 10 controls, deep learning-generated z-score maps with 3D T1 yielded accuracy 0.98 (not significant) but improved inter-rater agreement from κ=0.525 to 0.792.

• Diagnostic confidence improved for one rater with z-score maps vs baseline 3D T1 (adjusted p=0.017).
• Z-score maps were generated from 3T T1 MRI using stochastic cortical self-reconstruction (SCSR) and FreeSurfer.
• Three neuroradiologists rated cortical thickness or z-score maps, each with or without 3D T1, against standard 3D T1 reads.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.