Results Reference

DiffBench stores experiment-level and aggregated outputs in structured files so that results can be inspected, compared, or processed programmatically.

For an overview of the analysis workflow, see Analysis.

Analysis outputs

The analysis pipeline typically writes aggregated results under:

exp_outputs/summary/

global_metrics.parquet

Contains the fold-level metrics collected across experiments.

This table can be used when analyses require access to individual cross-validation folds rather than only experiment-level summaries.

summary_metrics.parquet

Contains metrics aggregated across folds for each experiment.

Typical summaries include the mean and standard deviation of the evaluation metrics.

comprehensive_table.parquet

Contains one row per experiment and combines:

• aggregated performance metrics;

• experiment identifiers;

• flattened experiment configuration values.

This table is useful for comparing experiment settings and filtering benchmark results.

Dataset reports

Files following the pattern:

<dataset>_report.txt

provide human-readable summaries for individual datasets, including experiment comparisons and best-performing runs.

Coverage table

coverage_table.txt

summarizes which combinations of datasets, models, targets, and tissue representations are available in the experiment results.

Coverage table encoding

Coverage cells use a compact two-character code describing the tissue representation and prediction target.

The first character represents the tissue:

g = gray matter
w = white matter

The second character represents the target:

g = gender
a = age
d = diagnosis

For example:

gg

represents gray matter with gender as the prediction target, while:

wa

represents white matter with age as the prediction target.

Plot outputs

Per-run plots are stored under:

exp_outputs/plots/<run_id>/

The exact visualisations depend on the experiment and the selected analysis configuration.