Fairboard: a quantitative framework for equity assessment of healthcare models

Résumé fourni par la source

# Model inferences for "Fairboard: a quantitative framework for equity assessment of healthcare models" This dataset contains the complete set of brain tumour segmentation inferences (NIfTI volumes) and voxel-wise performance metrics (CSV) for 18 open-source segmentation models evaluated across 648 glioma patients from two independent datasets: UCSF-PDGM (n = 501) and UPENN-GBM (n = 147), totalling 11,664 model inferences. These data underpin all analyses reported in Ruffle et al., "Fairboard: a quantitative framework for equity assessment of healthcare models." ## Contents The archive is organised by model, with each model directory containing: - `REFERENCE.txt` — full citation, DOI, and model notes - `UCSF-PDGM/inferences.csv` — per-patient performance metrics (501 rows, 71 columns) - `UCSF-PDGM/*.nii.gz` — predicted segmentation volumes (one per patient) - `UPENN-GBM/inferences.csv` — per-patient performance metrics (147 rows, 71 columns) - `UPENN-GBM/*.nii.gz` — predicted segmentation volumes (one per patient) ## Models (18) Ferreira et al. (2024), Myronenko et al. (2025), Maani et al. (2024), Luu and Park (2022), Swin UNETR (2022), DeepSCAN (2021), nnU-Net (2021), SANet (2021), Wang et al. (2021), H2NF-Net (2021), Zhao et al. (2020), Feng et al. (2020), No New-Net (2019), Marcinkiewicz et al. (2019), 3D-ESPNet (2019), Weninger et al. (2019), SegResNet (2019), and HD-GLIO (2019). Models span the BraTS 2018–2023 challenge series. Full citations are provided in each model's `REFERENCE.txt`. ## Performance metrics (per tumour compartment) Metrics are reported for four tumour sub-regions — whole tumour (WT), enhancing tumour (ET), peritumoral oedema (ED), and non-enhancing tumour/necrotic core (NET) — and include: Dice similarity coefficient, 95th-percentile Hausdorff distance (mm), intersection over union, sensitivity, specificity, precision, balanced accuracy, Matthews correlation coefficient, average surface distance (mm), normalised surface distance at 1 mm tolerance, volume similarity, and voxel-level confusion matrix counts (TP, FP, FN, TN). ## Segmentation volumes NIfTI files follow the BraTS labelling convention: 0 = background, 1 = necrotic tumour core (NET), 2 = peritumoral oedema (ED), 4 = enhancing tumour (ET). All volumes are 240 × 240 × 155 voxels at 1 mm isotropic resolution in the BraTS preprocessed coordinate space (SRI24 atlas). HD-GLIO segmentation volumes contain only labels 2 (ED) and 4 (ET), consistent with its 2-class design. ## Source datasets The original source MRI volumes, ground truth segmentation masks, and patient metadata are available from the respective dataset repositories: - UCSF-PDGM: Calabrese et al. (2022), "The University of California San Francisco Preoperative Diffuse Glioma MRI Dataset." *Radiology: Artificial Intelligence*, 4(6), e220058. https://doi.org/10.1148/ryai.220058 - UPENN-GBM: Bakas et al. (2022), "The University of Pennsylvania Glioblastoma (UPenn-GBM) Cohort." *Scientific Data*, 9, 453. https://doi.org/10.1038/s41597-022-01560-7 ## Associated publication Ruffle, J.K., Mohinta, S., Foulon, C., Zeina, M., Wang, Z., Brandner, S., Hyare, H. & Nachev, P. (2026). "Fairboard: a quantitative framework for equity assessment of healthcare models" *arXiv* preprint. ## Associated code Fairboard dashboard: https://fairboard.streamlit.app

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.