From a Clinical Mammography Archive to an AI-Ready Imaging Cohort: A Single-Center Implementation and Evidence Audit.
Authors
Affiliations (1)
Affiliations (1)
- Laboratory of Molecular and Surgical Research, Department of Research, Changhua Christian Hospital, No. 235, XuGuang Road, 8F., , Changhua, Taiwan. [email protected].
Abstract
Clinical imaging archives contain potentially valuable research data, but records retained for patient care do not automatically form an analysis-ready cohort. In this single-center study, we implemented and audited a metadata-driven workflow that separated patient-day indexing, image/report availability, study-defined bilateral four-view structure, quality control state, and pathology linkage. Among 13,203 indexed patient-day groups from January 2015 through December 2020, 5390 were DICOM-bearing, and 3169 of these met the metadata-defined bilateral four-view requirement (58.8%). The same requirement was met by 3168 of 5381 DICOM- and report-matched folders (58.9%), whereas 7813 report-linked groups lacked retained DICOM and were not assessable for image structure. Pathology was examined independently rather than as a nested endpoint; the strict-window diagnostic biopsy analysis identified 229 strict-window high-confidence groups. Thus, archive size, joint image/report availability, structural eligibility, and outcome evidence defined different analytical states. This implementation shows how established imaging-informatics principles can make cohort construction auditable within one institution; transferability requires independent external implementation and evaluation.