One-Shot Medical Landmark Detection

The success of deep learning methods relies on the availability a large number datasets with annotations; however, curating such is burdensome, especially for medical images. To relieve burden landmark detection task, we explore feasibility using only single annotated image and propose novel framework named Cascade Comparing to Detect (CC2D) one-shot detection. CC2D consists two stages: 1) Self-supervised (CC2D-SSL) 2) Training pseudo-labels (CC2D-TPL). CC2D-SSL captures consistent anatomical information in coarse-to-fine fashion by comparing cascade feature representations generates predictions training set. CC2D-TPL further improves performance new detector those predictions. effectiveness evaluated widely-used public dataset cephalometric detection, which achieves competitive accuracy 86.25.01% within 4.0 mm, comparable state-of-the-art semi-supervised lot more than one image. Our code available at https://github.com/ICT-MIRACLE-lab/Oneshot_landmark_detection.