Artificial intelligence in breast positioning and quality assurance in mammography

Abstract

Accurate breast positioning in mammography is essential for diagnostic image quality and quality assurance, yet image-evaluation systems such as PGMI (Perfect, Good, Moderate, Inadequate) are subjective, time-consuming and prone to inter-and intra observer variability. This dissertation aimed to train, test, and validate deep learning models for assessing breast positioning on medio-lateral oblique views using the posterior nipple line (PNL) criterion, and to evaluate radiographers’ perceptions of AI in breast positioning and quality assurance. A mediolateral-oblique subset of the VinDr-Mammo dataset (n=2,000) was matched by SOPInstanceUID to the deep-breast-positioning GitHub repository, and models were replicated. Two strategies were studied: (i) landmark regression (by replicating the U Net, Attention U-Net, CoordAtt U-Net and ResNeXt-50 models, and employing a novel HRNet), with Good/Bad labels derived post hoc via a deterministic PNL rule, and (ii) direct image-level classification (ResNeXt-50 replica, Optuna-tuned ResNeXt-50, ConvNeXt-Tiny, and EfficientNet-B3). The performance metrics for regression included per-landmark Euclidean error (mm) and pectoral-line angular error (°), while those for classification included macro-F1 and ROC-AUC on the test set. Results were reported as mean ±standard deviation across five seeds. In parallel, a prospective cross-sectional questionnaire was distributed amongst radiographers working in the mammography unit (n=9) at a local general public hospital in Malta. For regression, HRNet yielded the lowest landmark and angular errors and, via the PNL rule, the strongest derived classification (accuracy 94.20±1.04%; F1(Bad) 92.67±1.27%). For direct classification, ConvNeXt-Tiny provided the most balanced performance (macro-F1 82.64±2.07%; accuracy 83.40±2.22%), while EfficientNet-B3 was lower on macro-F1 (82.10±2.42%) but achieved the highest Sensitivity(Bad) (84.16±5.91%) and ROC-AUC (90.65±2.57%); both exceeded ResNeXt-50 baselines. Questionnaire response rate was 88.9%. PGMI was viewed as subjective (4.25/5) and time-consuming (3.75/5). Adoption enablers were workflow integration (n=6) and training (n=5); concerns were over-reliance (n=7), accountability (n=6) and reduced autonomy (n=5). Amongst the approaches evaluated, HRNet achieved the strongest landmark regression performance and consequently the best post-hoc PNL-derived Good/Bad grading, whereas for direct image-level classification, ConvNeXt-Tiny provided the most balanced overall performance, with EfficientNet-B3 achieving the highest ROC AUC and sensitivity for Bad cases. Questionnaire findings indicate that radiographers perceive practical value in AI support for positioning, particularly for improving consistency and enabling real-time feedback, while emphasising that adoption depends on training and workflow integration. However, external validation on independent datasets is required to confirm generalisable performance prior to prospective evaluation in clinical practice.