Dataset: Data quality biases normative models derived from fetal brain MRI

These are the data for the following article Thomas Sanchez*, Angeline Mihailov*, Gerard Martí-Juan, Nadine Girard, Aurélie Manchon, Mathieu Milh, Elisenda Eixarch, Vincent Dunet, Mériam Koob, Léo Pomar, Joanna Sichitiu, Miguel A. González Ballester, Oscar Camara, Gemma Piella, Meritxell Bach Cuadra**, Guillaume Auzias**; Data quality biases normative models derived from fetal brain MRI. Imaging Neuroscience 2026; doi: https://doi.org/10.1162/IMAG.a.1337 It is the companion to the code available on GitHub, and is structured as follows. Data and outputs structure <Region> below stands for one of the eight brain structures analyzed throughout the repository: eCSF_Ventricles_total, cGrey_Matter_Volume, White_Matter_Volume, Lateral_Ventricles_Volume, Cerebellum_Volume, Basal_Ganglia_Volume, Brainstem_Volume, Thalamus_Volume. data/ data/ ├── raw_data/ │ └── finalized_df_qcglobal_and_z.csv # Raw unharmonised CSV file. ├── harmonization/ # ComBat-harmonized datasets, full cohort (incl. dHCP) │ ├── harmonized_data_all.csv # Data harmonized using data from all subjects │ ├── harmonized_data_poor_plus.csv # Data harmonized using "PoorPlus" quality: qcglobal >= 0.95 │ ├── harmonized_data_accept_plus.csv # Data harmonized using "AcceptPlus" quality: qcglobal >= 1.95 │ ├── harmonized_data_ground_truth.csv # Data harmonized using "GroundTruth" quality: qcglobal >= 2.70 │ └── iterative/ # Harmonization using an iterative removal of data according to quality. │ └── harmonized_data_{0,5,10,...,600}.csv # Same cohort with the N lowest-quality subjects removed (121 files, step 5) └── harmonization_nodhcp/ # Same as harmonization, with dHCP cohort left-out The four QC tiers (All / PoorPlus / AcceptPlus / GroundTruth) and their qcglobal thresholds are defined in the paper and in the GitHub repository notebooks/1_harmonize_data.ipynb. outputs_indiv/ Each subfolder is the output of one script in the scripts/ provided in the GitHub repository . Within a subfolder, files come in matched pairs per brain structure: an analysis/curves` CSV holding the bootstrapped centile trajectories, and a matching `*_indices_*` CSV recording which subject indices were resampled in each bootstrap round. outputs_indiv/ ├── base_analysis/ # scripts/1_run_main_analysis.r — main GAMLSS fits across all QC tiers │ ├── main_grid.svg # Combined centile-curve figure across structures/tiers │ ├── original_analysis_<Region>.csv # Bootstrapped centile curves (age, centile, qc tier, value) │ └── original_indices_<Region>.csv # Bootstrap resample indices ├── base_analysis_nodhcp/ # Same as base_analysis/, no-dHCP cohort robustness check │ ├── main_grid.svg │ ├── original_analysis_<Region>.csv │ └── original_indices_<Region>.csv ├── base_analysis_fixed_harmo/ # scripts/2_run_fixed_harmonization_ablation.r — fixed-harmonization ablation │ ├── original_analysis_<Region>.csv │ └── original_indices_<Region>.csv ├── bootstrap_qc/ # scripts/3_run_bootstrapping_experiment.r — QC tier x sample-size bootstrap │ ├── bootstrapped_curves_<Region>.csv # Centile curves per QC tier, target quality (lambda) and sample size │ ├── bootstrapped_indices_<Region>.csv # Bootstrap resample indices (incl. realized QC, lambda) │ └── table_bootstrapping_results.csv # Regression of centile value on target quality & sample size, per region/centile └── iterative_removal/ # scripts/supp_run_iterative_removal.r — progressive worst-quality-subject removal ├── iterative_analysis_<Region>.csv # Centile curves vs. number/quality of subjects removed (nsamp, min_qc, avg_qc) └── iterative_indices_<Region>.csv # Bootstrap resample indices per removal step