Supported Downstream Tasks
NeuroSTORM supports 5 core downstream tasks for fMRI analysis, each designed to evaluate different aspects of neural representation learning and model generalization. Select a compatible model input and dataset preset using the model support and dataset support tables. The common supervised runner covers all twelve models.
1. Age and Gender Prediction
Task Description:
This task assesses the ability of models to predict basic demographic variables—chronological age and biological sex—using only resting-state fMRI (rsfMRI) sequences as input. The input is a preprocessed fMRI representation for each subject, using the selected model's voxel, ROI, FC, or graph format. The output is either a continuous age value (for regression) or a categorical label (male/female, for classification). Age and sex are fundamental variables correlated with brain structure and function, and accurate prediction indicates that brain representations capture meaningful demographic information. This task is widely used as a baseline for evaluating the generalizability and biological relevance of neural representations in fMRI analysis.
🔧 How to use in NeuroSTORM
-
Sex classification:
--downstream_task_id 1 --downstream_task_type classification --task_name sex -
Age prediction (regression):
--downstream_task_id 1 --downstream_task_type regression --task_name age
2. Phenotype Prediction
Task Description:
This task involves predicting quantitative or categorical phenotypic scores (such as cognitive, behavioral, or clinical measurements) from fMRI data. The input is a preprocessed rsfMRI or tfMRI representation in the selected model's format, and the output is a continuous score for the target phenotype. Example outputs include MMSE scores, PANSS scores, DASS measures, and other clinical or cognitive test results. This task provides a direct evaluation of how well neural representations capture individual differences in brain function related to cognition, emotion, or disease traits, and is critical for developing clinically useful neuroimaging biomarkers.
🔧 How to use in NeuroSTORM
-
Phenotype regression:
--downstream_task_id 2 --downstream_task_type regression --task_name your_score_name
Replaceyour_score_namewith the phenotype to predict (e.g.,MMSE,PANSS_Positive).
3. Disease Diagnosis
Task Description:
This task requires the model to assign each subject to a diagnostic category (e.g., healthy control, ADHD, schizophrenia, autism) based on their resting-state or task-based fMRI data. The input is a preprocessed fMRI representation in the selected model's format for each subject. The output is a categorical disease label. Disease diagnosis tasks are central for translational neuroimaging, as they test the capacity of models to extract pathological signatures from brain activity and are directly relevant to clinical decision support and biomarker discovery.
- ADHD-200: ADHD vs. Control
- ABIDE I and ABIDE II: Autism vs. Control
- COBRE: Schizophrenia vs. Control
- HCP-EP: Early Psychosis vs. Control
- UCLA: Multi-disorder classification (Schizophrenia, Bipolar, ADHD, Control)
- TransDiag: Diagnosis labels from the prepared transdiagnostic metadata
🔧 How to use in NeuroSTORM
-
Disease classification:
--downstream_task_id 3 --downstream_task_type classification --task_name diagnosis
4. fMRI Re-identification (ReID)
Task Description:
This task evaluates subject-specific scan embeddings for identity matching, using the volume or ROI/FC representation required by the selected retrieval model. Under a closed-set protocol, each query fMRI sequence is matched against a non-overlapping gallery of fMRI sequences to retrieve the correct identity. This task tests whether the learned representation preserves individual-specific functional signatures and supports reliable similarity search in high-dimensional fMRI space.
Experimental Procedure:
- Dataset and Protocol: Use the held-out HCP-YA split and the supplied fixed query/gallery protocol from the dedicated Task4 evaluator.
- Embedding Extraction: Apply the evaluator's model-specific input transformation and scan feature aggregation to obtain one L2-normalized embedding per scan.
- Retrieval: For each query embedding, perform exhaustive nearest-neighbor search in the gallery embedding space (standard NN in the normalized feature space).
- Evaluation Metrics: Report Rank-1 Accuracy and mean Average Precision (mAP) using the supplied query/gallery protocol consistently across models.
🔧 How to use in NeuroSTORM
-
fMRI re-identification (ReID):
python scripts/task4/evaluate_reid.py --help
Use the dedicated evaluation instructions for NeuroSTORM, SwiFT, BrainLM, TFF, BNT, Com-BrainTF, LG-GNN, and BrainGNN.
5. Task-based fMRI State Classification
Task Description:
This task involves classifying which cognitive state or task condition a subject is in, based on their task-based fMRI (tfMRI) sequence. The input is a preprocessed tfMRI sequence corresponding to a specific cognitive experiment (e.g., language, emotion, gambling task). The output is a label indicating the task condition or cognitive state. Accurate state classification demonstrates that the model captures functionally relevant brain activation patterns, and this task is a key benchmark for evaluating generalization and sensitivity to cognitive manipulations in fMRI analysis.
HCPTASK: Emotion, Gambling, Language, Motor, Relational, Social, Working MemoryMOVIE: Movie/group labels from the prepared dataset metadata and task configuration
🔧 How to use in NeuroSTORM
-
Task state classification:
bash scripts/run_experiment.sh --model band --dataset hcptask --task task5 --mode train_scratch
Select another shared backbone with its matching input representation and task configuration.
Task Selection Guide
| Task | Type | Input Type | Best For |
|---|---|---|---|
| Age & Gender Prediction | Regression / Classification | rsfMRI | Evaluating representation quality and generalization |
| Phenotype Prediction | Regression | rsfMRI / tfMRI | Biomarker discovery and clinical translation |
| Disease Diagnosis | Classification | rsfMRI | Clinical decision support and disease understanding |
| fMRI ReID | Classification / Retrieval | rsfMRI | Testing individual-specific neural signatures |
| Task State Classification | Classification | tfMRI | Evaluating sensitivity to task-driven activation |
Custom Tasks
NeuroSTORM supports creating custom downstream tasks by:
-
Defining the dataset label format in the
make_subject_dictfunction from data_module.py -
Setting
--downstream_task_id,--downstream_task_type, and--task_namein the experiment configuration - Choosing a classification or regression head. For custom tasks, add a new head definition in models/heads