Reproduction Package for Auditing Tiny AgNOR Satellite Segmentation: Reproducible Metric Failures and Group-Aware Evaluation Under Point Supervision

作者：Wenbin Zhang, H Z Chen, Yang Li, G Chang, X Zhao, Xiaoling Yang, Hao Wang, Binhang Huang · 发表于：Zenodo (CERN European Organization for Nuclear Research) · 年份：2026 · DOI：10.5281/zenodo.21537346 · 研究领域：Computer science、Data mining、Real-time computing、Algorithm

This reproduction package supports the study “Auditing Tiny AgNOR Satellite Segmentation: Reproducible Metric Failures and Group-Aware Evaluation Under Point Supervision.” It contains executable counterexamples and independent reference implementations for auditing the released CCAgT evaluation code, standard class-specific evaluation of 331 released Version 1 predictions, 5,000 seeded randomized metric-property tests, code-line evidence, and figure-generation scripts. It also contains the frozen configurations, group-disjoint five-fold splits, three-seed results, case- and group-level outputs, training logs, and final model weights for experiments on 9,339 Version 2 fields from 15 released patient/slide groups. The original CCAgT RGB images are not redistributed and must be obtained from the cited Mendeley Data records. The Version 2 satellite masks are point-derived raster proxies and are not interpreted as manually delineated boundaries. Some executable fields and source-code identifiers retain the historical terms “patient_equal” or “LOPO.” In this archive, these identifiers refer to equal weighting or holding out of a released prefix group; they do not independently verify that every prefix represents a different patient. This work was supported by the Inner Mongolia Autonomous Region Science and Technology Plan, Project No. 2023YFSH0052.