nikhil153 / brain-diff

brain-diff

Goals:

1. Brainage prediction with two timepoints
2. Brainage Biomarker for AD and PD

Related work

- Peng H, Gong W, Beckmann CF, Vedaldi A, Smith SM. Accurate brain age prediction with lightweight deep neural networks. Med Image Anal. 2021. [Paper](https://doi.org/10.1016/j.media.2020.101871), [Code](https://github.com/ha-ha-ha-han/UKBiobank_deep_pretrain)
- Jonsson, B.A., Bjornsdottir, G., Thorgeirsson, T.E. et al. Brain age prediction using deep learning uncovers associated sequence variants. Nat Commun 10, 5409 (2019). [Paper](https://www.nature.com/articles/s41467-019-13163-9), Code on request.
- Baecker L, Garcia-Dias R, Vieira S, Scarpazza C, Mechelli A. Machine learning for brain age prediction: Introduction to methods and clinical applications. EBioMedicine. 2021 [Paper](https://pubmed.ncbi.nlm.nih.gov/34614461/)
- Leonardsen EH, Peng H, ... Wang Y. Deep neural networks learn general and clinically relevant representations of the ageing brain. Neuroimage. 2022. [Paper](https://pubmed.ncbi.nlm.nih.gov/35462035/)

Datasets

- UKBB
- ADNI
- PPMI

UKB data wrangling

- Copy files from squashfs on Beluga
- Organize them in psudo-bids

for i in `ls | grep sub- | grep -v json`; do 
    mkdir -p ../`echo $i | cut -d "_" -f1`/ses-2/anat; 
    mv `echo $i | cut -d "_" -f1`* ../`echo $i | cut -d "_" -f1`/ses-2/anat/;  
done

ADNI data wrangling

- use src/generate_adni_bids.py

MR preprocessing

- fmriprep anat workflow
    - template spaces: `MNI152NLin2009cSym_res-1`, `MNI152NLin6Sym_res-1`, `MNI152Lin_res-1`
- freesufer 6.0.1
    - DKT (n_rois: 31x2) cortical thickness and ASEG volumes

Experiments: model training with controls (UKBB)

- input_visit --> output_visit
    1. Baseline --> Baseline
    2. Baseline + Followup --> Baseline
    3. Baseline + Followup --> Baseline + Followup
- features + models
    1. DKT (Ridge, RF)
    2. T1w normalized to the MNI template(s) (SFCN, LSN)
- nulls
    1. Median (+2) age prediction
- perf metrics
    1. mean abs error
    2. pearson's r
    3. temporal consistency 

Experiments: model biases on control cohorts

- age vs brainage_error bias (effectiveness of linear correction)
- short vs long visit_delta: UKB longterm cohort (FU - BL > 3yr)
- study+scanner variation: ADNI, PPMI control cohorts

Experiments: brainage gap

- Single number: Disease stages vs study-specific controls vs long_visit UKBB vs short_visit ukbb
    - Note even with two visits only BL brainage value is likely to be useful. 
- DeepNet representations: Richer constellation / clusterring of subjects from model embeddings 

---

Simulations:

- Simple interactive runs: notebooks/7_brain_diff_sim.ipynb
- Batch runs: src/run_simul.py

SFCN depoyment on new data:

- src/run_SFCN.py