Research · 2022 · DeepPET-3D · Research at KJ Somaiya College of Engineering
Alzheimer's Disease Detection via PET Imaging
Can a scan catch the stage before dementia?
3D convolutional neural network for early diagnosis of Alzheimer's disease from 18F-FDG PET brain scans, separating cognitively normal, mild cognitive impairment and Alzheimer's at 92.31% accuracy, and up to 96.91% pairwise.
Problem
Alzheimer's is easiest to slow before it becomes dementia, and the stage before that, mild cognitive impairment, looks closest to a healthy brain. Telling the three apart from an FDG-PET volume is the task.
Approach
- 18F-FDG PET 3D axial scans from the ADNI repository, labelled cognitively normal (CN), mild cognitive impairment (MCI) or Alzheimer's disease (AD).
- Preprocessing on the DICOM volumes before augmentation: thresholding, voxel normalisation and volume reduction.
- A 3D convolutional neural network trained for one three-way experiment and three pairwise ones.
Stack
- Python
- TensorFlow
- Keras
- OpenCV
- W&B
Outcome
- 92.31% accuracy across the three classes
- AD vs CN 94.79%, CN vs MCI 93.28%, MCI vs AD 96.91%
- The closest pair, normal against mild impairment, is the hardest to separate