Lightweight semi-supervised coarse-to-fine landmark matching in chest CT using a Siamese network

Anatomical landmark localization is a key component in many medical image analysis workflows, including image registration, atlas-based segmentation, surgical planning, and longitudinal disease monitoring. Accurate correspondence across scans is challenging due to the high cost of manual labeling and variability among annotators. We present a semi-supervised Siamese network–based method for landmark matching in chest CT, designed to operate efficiently with minimal manual annotation. The approach uses sparsely sampled image descriptors to capture rich anatomical context while keeping computational cost low. Training labels are generated automatically via a point matching algorithm applied to a small set of manually labeled reference scans, enabling the model to learn anatomical embeddings that generalize across subjects and landmark types. Experimental results show that incorporating spatial attention into a lightweight backbone improves coarse displacement prediction, while a hierarchical coarse-to-fine strategy further reduces localization error. The proposed method achieves millisecond-level inference per query, making it suitable for real-time navigation and large-scale processing. These findings demonstrate the potential of combining sparse descriptor sampling, semi-supervised learning, and hierarchical refinement for efficient and accurate landmark matching in large-displacement scenarios, and provide a framework that can be extended to other imaging modalities and anatomical regions.