EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Lesion-Location Supervision for Dual-Resolution Mammography Classification: A Patient-Grouped Evaluation on Mini-MIAS

Muhammad Abuarrah, Yogeesh N, Afshin Soflaei Shahrbabak

Qubahan Journal of Medical Sciences · 12.11.2026 · 3 · 76-90

DOI: 10.48161/qjms.v3a92 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Qubahan Journal of Medical Sciences, online erschienen am 12.11.2026. Verfasst von Muhammad Abuarrah und 2 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study evaluated whether lesion-location supervision improves benign-versus-malignant classification and lesion-center coverage in full mammograms. The analysis included 115 Mini-MIAS mammograms from 108 patients, comprising 63 benign and 52 malignant images. Two dual-resolution models combined a downsampled whole-mammogram view with 25 overlapping high-resolution tiles using a shared frozen ResNet-18 encoder. One model used lesion centers to guide tile weighting during training. Both were predicted from pixels alone on the same held-out patient folds. In the reference run, pooled area under the receiver operating characteristic curve (ROC-AUC) was 0.493 with location supervision and 0.491 with image labels alone (paired difference 0.002, 95% patient-bootstrap confidence interval −0.006 to 0.011). Across five training runs on fixed partitions, supervised ROC-AUC ranged from 0.480 to 0.511. Among 111 mammograms with usable coordinates, top-tile lesion-center coverage increased from 6.3% to 23.4% (gain 17.1 percentage points, 95% confidence interval 9.1 to 25.5). Uniform random and fixed central tile selection yielded 6.5% and 39.6% coverage, respectively. On the coordinate-eligible subset, reference ROC-AUC was 0.488 for the annotation-centered crop classifier and 0.508 for the supervised full-image model. Location supervision improved tile coverage, while full-image malignancy discrimination stayed near chance under the evaluated protocol.

Zitieren

APA

Abuarrah, M., N, Y., & Shahrbabak, A. S. (2026). Lesion-Location Supervision for Dual-Resolution Mammography Classification: A Patient-Grouped Evaluation on Mini-MIAS. Qubahan Journal of Medical Sciences, 3, 76-90. https://doi.org/10.48161/qjms.v3a92

Vancouver

Abuarrah M, N Y, Shahrbabak AS. Lesion-Location Supervision for Dual-Resolution Mammography Classification: A Patient-Grouped Evaluation on Mini-MIAS. Qubahan Journal of Medical Sciences. 2026;3:76-90. doi:10.48161/qjms.v3a92

BibTeX
@article{Abuarrah2026,
  author = {Abuarrah, Muhammad and N, Yogeesh and Shahrbabak, Afshin Soflaei},
  title = {Lesion-Location Supervision for Dual-Resolution Mammography Classification: A Patient-Grouped Evaluation on Mini-MIAS},
  journal = {Qubahan Journal of Medical Sciences},
  year = {2026},
  volume = {3},
  pages = {76--90},
  doi = {10.48161/qjms.v3a92}
}
Kostenlos starten →