Knee osteoarthritis affects millions worldwide and is often accompanied by lower limb malalignment. Clinicians measure leg alignment on X-rays to support diagnosis, surgical planning, and verification of prosthesis placement. This evaluation relies on manual landmark identification, a time-consuming process prone to interobserver variability. Existing automated solutions predominantly employ regression-based deep learning techniques. In this thesis, we explored an alternative strategy for landmark localization: image segmentation with circular masks centered at the landmark location, and evaluated its value for assessing lower-limb alignment. The first contribution introduces a deep learning topology termed segmentation-guided coordinate regression for landmark localization. While segmentation-based methods alone yield low average distance errors, the presence of false positives and missed detections limits clinical applicability. To solve this, we integrate a segmentation network with a coordinate regression branch, jointly trained end-to-end. This hybrid model achieves higher localization accuracy than standard coordinate regression and greater robustness than standalone segmentation. The second contribution focuses on refining segmentation-based approaches to reduce architectural complexity. We systematically evaluated network architectures, post-processing strategies for coordinate estimation, and segmentation mask sizes. A convolutional model trained with masks of radius 15 pixels, coupled with an adaptive threshold-based centroid extraction algorithm, achieved superior accuracy compared to segmentation-guided, heatmap-based, and coordinate-regression methods. Notably, this optimized approach improved performance in knee phenotype classification. The third contribution addresses the detection of inaccurate predictions that may compromise clinical decision-making. We propose a contrastively pretrained Siamese network that compares patches centered on predicted landmark positions with reference embeddings to assess similarity. The method was evaluated for identifying inaccurate femoral trochlear notch estimates. Results show that the model accurately detects deviations exceeding 2.0 mm from the ground truth and estimates error magnitude. Overall, the approach outperforms baseline regression and classification methods, proving generalization and effectiveness as a quality control mechanism. In conclusion, this thesis advances the field of landmark localization and demonstrates its clinical relevance for automated assessment of lower limb alignment. Landmark segmentation emerged as a highly competitive paradigm, setting the way for future research. Beyond accuracy, our contributions address robustness and failure identification, two aspects often overlooked but vital for future clinical deployment.
Amador Sanchez, S 2026, 'Advancing landmark localization through deep segmentation for reliable malalignment assessment in lower limb radiographs', Vrije Universiteit Brussel.
Amador Sanchez, S. (2026). Advancing landmark localization through deep segmentation for reliable malalignment assessment in lower limb radiographs. [PhD Thesis, Vrije Universiteit Brussel]. Crazy Copy Center Productions.
@phdthesis{6d432f23132e4450a85408e581fa9958,
title = "Advancing landmark localization through deep segmentation for reliable malalignment assessment in lower limb radiographs",
abstract = "Knee osteoarthritis affects millions worldwide and is often accompanied by lower limb malalignment. Clinicians measure leg alignment on X-rays to support diagnosis, surgical planning, and verification of prosthesis placement. This evaluation relies on manual landmark identification, a time-consuming process prone to interobserver variability. Existing automated solutions predominantly employ regression-based deep learning techniques. In this thesis, we explored an alternative strategy for landmark localization: image segmentation with circular masks centered at the landmark location, and evaluated its value for assessing lower-limb alignment. The first contribution introduces a deep learning topology termed segmentation-guided coordinate regression for landmark localization. While segmentation-based methods alone yield low average distance errors, the presence of false positives and missed detections limits clinical applicability. To solve this, we integrate a segmentation network with a coordinate regression branch, jointly trained end-to-end. This hybrid model achieves higher localization accuracy than standard coordinate regression and greater robustness than standalone segmentation. The second contribution focuses on refining segmentation-based approaches to reduce architectural complexity. We systematically evaluated network architectures, post-processing strategies for coordinate estimation, and segmentation mask sizes. A convolutional model trained with masks of radius 15 pixels, coupled with an adaptive threshold-based centroid extraction algorithm, achieved superior accuracy compared to segmentation-guided, heatmap-based, and coordinate-regression methods. Notably, this optimized approach improved performance in knee phenotype classification. The third contribution addresses the detection of inaccurate predictions that may compromise clinical decision-making. We propose a contrastively pretrained Siamese network that compares patches centered on predicted landmark positions with reference embeddings to assess similarity. The method was evaluated for identifying inaccurate femoral trochlear notch estimates. Results show that the model accurately detects deviations exceeding 2.0 mm from the ground truth and estimates error magnitude. Overall, the approach outperforms baseline regression and classification methods, proving generalization and effectiveness as a quality control mechanism. In conclusion, this thesis advances the field of landmark localization and demonstrates its clinical relevance for automated assessment of lower limb alignment. Landmark segmentation emerged as a highly competitive paradigm, setting the way for future research. Beyond accuracy, our contributions address robustness and failure identification, two aspects often overlooked but vital for future clinical deployment.",
author = "\{Amador Sanchez\}, Sebastian",
year = "2026",
language = "English",
isbn = "9789493461536",
publisher = "Crazy Copy Center Productions",
address = "Belgium",
school = "Vrije Universiteit Brussel",
}