Answers are grounded in this paper and cite their sources.
Automatic landmark detection is increasingly adopted in medical imaging, yet erroneous predictions remain common, requiring expert review. We propose a Siamese network-based quality-control framework that identifies inaccurate landmark detections and estimates the magnitude of their localization errors. Using a contrastive loss, the network learns a shared embedding space from image-patch pairs sampled near and far from the ground-truth position. During inference, test-patch embeddings are compared to K nearest reference embeddings using a similarity metric to determine landmark correctness and estimate distance to the ground-truth position. We evaluated the approach on full-leg radiographs targeting the femoral trochlear notch, optimized key hyperparameters, and compared two inference strategies: a library of embeddings from patches sampled at known distances from the ground truth and reference embeddings derived from landmark-centered patches. The clinical relevance of localization errors was assessed using directional landmark perturbations and a sensitivity analysis of phenotype assignment. The library-based strategy provided a balanced trade-off between distance estimation and classification, achieving a mean absolute error of 0.94 millimeters (mm) and an F 1 -score of 0.84 for distinguishing 0–2.0 mm from 2.0–4.0 mm patches. On an independent dataset, it correctly identified two major localization errors with a false-positive rate of 0.06. Medial and lateral displacements of 2.0 mm changed the assigned phenotype class in more than 10% of cases. These results highlight the potential of similarity-based Siamese embeddings for automated quality control in landmark detection and underscore the clinical relevance of landmark localization for downstream alignment assessment, supporting further evaluation across other anatomical landmarks and imaging applications.
No takes yet. Share an insight, caveat, or question.
Sánchez et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: