Landmark Regression with Attention U-Net for Assessing Breast Positioning Quality in MLO Mammography View.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, Sultan 2. Abdulhamid Han Training and Research Hospital, Istanbul 34668, Turkey.
- Biomedical Engineering Graduate Program, Istanbul Technical University, Istanbul 34469, Turkey.
- Department of Radiology, School of Medicine, University of Wisconsin, Madison, WI 53792, USA.
- Department of Radiology, Istanbul Training and Research Hospital, University of Health Sciences, Istanbul 34098, Turkey.
- Department of Radiology, Sisli Hamidiye Etfal Training and Research Hospital, Istanbul 34371, Turkey.
- Department of Radiology, School of Medicine, Acibadem Mehmet Ali Aydinlar University, Istanbul 34684, Turkey.
- Department of Computer Engineering, Istanbul Technical University, Istanbul 34469, Turkey.
Abstract
<b>Background/Objectives:</b> This study developed and evaluated a novel Attention U-Net regression model to assess MLO mammography positioning quality, comparing it against a plain U-Net regression baseline (architectural ablation) and a ResNeXt50 classification baseline. <b>Methods:</b> We curated 1000 patient mammograms (2000 MLO images) from the public VinDr-Mammo dataset, with pectoral muscle line and nipple positions annotated by an expert breast radiologist. Three deep learning models were compared under a 10-fold stratified cross-validation protocol. Statistical significance was assessed with paired Wilcoxon signed-rank tests, McNemar's test, Cochran's Q, Wilson score confidence intervals, bootstrap confidence intervals, Holm-Bonferroni correction, and paired Cohen's d effect sizes. <b>Results:</b> Against the automated PNL reference, the Attention U-Net achieved 85.5% accuracy (Wilson 95% CI [83.9%, 87.0%]), 80.2% sensitivity, and 88.7% specificity, outperforming both the plain U-Net (71.8% accuracy) and ResNeXt50 (73.7% accuracy). (Cochran's Q <i>p</i> < 0.001; pairwise McNemar <i>p</i> < 0.001). For landmark localization, the Attention U-Net produced significantly lower errors across all five evaluated landmark and geometric measures (nipple 3.44 mm vs. 7.02 mm; perpendicular intersection 5.93 mm vs. 9.81 mm; all comparisons remained significant after Holm-Bonferroni correction). <b>Conclusions:</b> In this single-dataset, single-view study, an Attention U-Net landmark-regression model provided explicit landmark and PNL outputs for quantitative MLO mammography positioning assessment. The model showed improved accuracy and sensitivity compared with a plain U-Net regression baseline and a ResNeXt50 classification baseline.