MODD: Multicenter one-shot medical landmark detection and denoising.

Objective: One-shot landmark detection on lateral cephalometric X-ray images has advantages for label-efficient training. As clinical datasets are typically small and do not generalize well to variations in equipment or imaging environments. When using datasets from multiple studies (multi-center data) for joint training, the imbalance in the number of images and the differences in image properties can lead to poor generalization performance. In this work, we aim to propose a method that leverages more data from multiple experiments to improve the accuracy of a single experiment. Approach. To address these challenges, we propose a Multicenter One-shot landmark Detection and Denoising framework (MODD). It incorporates a self-supervised one-shot mapping based on multicenter template transformation and the pseudo-label denoising module. Label denoising is used to reduce the impact of inaccurate pseudo-labels on the algorithm, focusing on label quality rather than signal noise in X-ray images. For denoising module, we propose the shuffled dynamic sample selection and contrastive correction of multicenter pseudo labels. These two components together enable more accurate one-shot landmarks detection on lateral cephalometric X-ray images in multicenter scenarios. Main results. Experiments are conducted with the publicly available multicenter cephalometric X-ray datasets. MODD achieves a landmark detection accuracy of 79.27\% within a 4.0 mm range and the mean radial error of 2.94 mm, demonstrating satisfactory performance compared to state-of-the-art methods. Significance. This study expands the application of medical landmark detection to the multi-center one-shot filed and demonstrates the potential of the MODD architecture.