Segmentation errors that track the disease: code, ground-truth masks and result tables

Code, ground-truth instance masks and result tables for the study "Segmentation errors that track the disease: a globally scaled seed threshold biases the sickle fraction in automated blood film analysis". We audit the erythrocyte segmentation released with a published sickle cell analysis pipeline against instance-level ground truth on 25 thin blood film fields (3,762 cells) and a batch-matched cohort of 24 slides (4,070 cells) from the public collection of Manescu et al. (2020), https://doi.org/10.5522/04/12407567. The released method recovers 55.8% of erythrocytes at IoU 0.5, recovers fewer cells from sickle cell films than from batch-matched controls (43.5% vs 62.3%), and under-reports the sickle fraction by about 17% in relative terms. The disparity is traced to a seed threshold scaled to the field-wide maximum of the distance transform; extended-maxima seeding reduces it from 18.8 to 4.6 points. Contents: gt_masks.zip (primary-cohort ground-truth masks), gt_masks_bm.zip (batch-matched cohort masks), paper2_results_final.zip (per-cell, per-field and per-slide result tables), sickle-segmentation-audit-main.zip (analysis notebook and figure code; also at https://github.com/Jenifer87/sickle-segmentation-audit).