GeoMed-M: A Benchmark-Centric Audit of Geographic Coverage in Public Medical-Imaging Datasets

Public medical-imaging benchmarks shape where models are developed and tested, yet the geography of their source data is rarely audited. We introduce GeoMed-M, a source-linked catalogue of 59 public dataset releases across nine imaging families. After removing five derivative or superseded releases, 44 of 54 primary releases had explicit country provenance. Only two came from Organisation of Islamic Cooperation (OIC) countries: BUSI (Egypt) and COVID-CT-MD (Iran). Together they contributed 905 of 435,669 patients (0.208\%) among releases with supported patient counts. The three largest source countries contributed 88.6\% of single-country patients, and seven of nine imaging families contained no OIC-country release. A separate census of 667 records from TCIA and PhysioNet reached the same conclusion:only 2 of 198 eligible records (1.01\%) explicitly reported collection in an OIC country. Sensitivity analyses, MedMNIST, and income-stratified results also showed sparse coverage. Reporting was uneven: only 42.6\% of releases documented collection periods and 25.9\% established patient-disjoint splits. GeoMed-M measures the visibility of health systems in public data infrastructure; country is used as collection context, never as a proxy for patient identity or model performance.