KAN-OSD: Dino-Based Encoder and KAN-Based Decoders with Dual Contrastive Learning for One-Shot Anatomical Landmark Detection

One-shot anatomical landmark detection seeks accurate localization from a single labeled template. To address this data scarcity issue, an effective strategy is to leverage the strong generalization of a ViT-based foundation-model encoder, particularly its out-of-domain knowledge. We propose KANOSD, which freezes the encoder and trains only a lightweight prediction head with distance-aware similarity and contrastive objectives. Conventional convolutional decoders are replaced by KolmogorovArnold Networks (KAN) to model a continuous, high-resolution similarity field conditioned jointly on foundation-model features and image coordinates. To further enhance discriminability under limited supervision, we introduce metric-aware contrastive learning, which uses multiscale and multi-view augmentations of the template image as network inputs. Contrastive learning across global and local features strengthens context consistency for each landmark, while diverse augmentation views help reduce overfitting and improve generalization. KAN-OSD maintains strong label and computational efficiency without requiring any unlabeled images, while achieving lower mean radial error and higher successful detection rates on Head and Hand X-ray benchmarks.