Question
Question Posted 03/15/25:
Dear ADNI Experts,
I am working on a project using VGG and MobileNet for Alzheimer's classification. Since these models require 2D inputs, I need to extract the most relevant 2D slices from NIfTI (.nii) files.
What is the best approach to select meaningful slices without losing important information? I prefer not to modify the models for 3D input, as I want to keep them lightweight.
I’d appreciate any guidance or recommended techniques.
Best regards,
Dear ADNI Experts,
I am working on a project using VGG and MobileNet for Alzheimer's classification. Since these models require 2D inputs, I need to extract the most relevant 2D slices from NIfTI (.nii) files.
What is the best approach to select meaningful slices without losing important information? I prefer not to modify the models for 3D input, as I want to keep them lightweight.
I’d appreciate any guidance or recommended techniques.
Best regards,
Response posted 03/17/25 by Jeff Gunter:
Feature selection is a key element in ML/AI. I would recommend starting with a literature search to identify what anatomic regions typically differentiate AD from normal aging. Once you have that list (it will likely include temporal and parietal lobe regions), consult an atlas to identify likely slices in a standard coordinate system. Then you'll need to propagate those slice locations into the coordinate system of each input image (or resample each input image into standard coordinate space). N.B. individual anatomy varies widely and patient positioning within the imaged field of view varies.
