Side-level versus patient-level evaluation in four-view mammography classification: a comprehensive benchmark on the CMMD dataset

Description

Mohamed A. Mabrok wrote a 2026 benchmark study on how breast X-ray classification results are reported. Deep learning papers on mammography often report scores per breast side or per patient, and that choice, he argues, makes results across studies hard to compare fairly.

Using 706 patients with four-view mammograms from a Chinese database, Mabrok tested six model designs crossed with three ways of combining views, under patient-level cross-checking. Side-level scores beat patient-level scores by an average of 17.5 percentage points, far more than the gaps between model designs, which were under three points. A large transformer model trailed every simpler network despite having many more parameters. The lesson is that reporting level drives headline performance.

Catalog record

Source: https://europepmc.org/article/MED/PMC13271282