These experiments utilize the LA5c Study from the Preprocessed Consortium for Neuropsychiatric Phenomics dataset. 265 participants completed extensive psychometric and neuroimaging examinations. A relatively simple modeling task was chosen: predict participants' questionnaire answers given their T1w MRI.
TODO: pick interesting questions we want to model with T1w
TODO: show a case where overtly visible differences in brain structure were exploited to infer correct answers
The dataset comes with hundreds of self-report answers for an amalgam of questionnaires. Participants' answers are modeled by the structural T1-weighted MRI of each participant. TODO: explain process behind selecting questions
TODO: show which questions were selected
| File | Input Size (CxDxHxW) | Notes |
|---|---|---|
| classification/bilingual.yaml | 1x176x256x256 | Model bilingual Y/N in terms of T1w |
| classification/bilingual_hparams.yaml | 1x176x256x256 | Hyperparameter search for bilingual.yaml |
| File | Notes |
|---|---|
| src/classification.py | Base classification experiment |
| src/dataset/la5c.py | LA5c dataset class |
| src/models/resnet_classifier3d.py | 3D ResNet classifier model |
LA5c is released under the Creative Commons Zero (CC0) license.
Apache 2.0 / MIT dual-license. Please contact me if this is somehow not permissive enough and we'll add whatever free license is necessary for your project.