In lateral spine landmark detection,the previous heatmap regression methods have difficulty in distinguishing landmarks on different vertebrae due to the influence of organ occlusion and are prone to landmark and vertebrae matching errors.To solve this problem,we propose a new one-stage lateral spine landmark detection method,which simultaneously predicts the landmark heatmap and landmark matching clue(vertebra center heatmap and landmark offset),and uses the matching clue to match the landmarks with the corresponding vertebra.In order to improve the matching effect,we propose the geometry-aware feature aggregator module,which can extract the landmark features on the vertebra to enhance the feature representation of the vertebra center.We also use a weighted loss function to alleviate the imbalance of positive and negative samples in the landmark and the vertebra center heatmaps.Experimental results show that the average detection error of the proposed method is 8.84,which has 36%improvement in accuracy compared to the method with the second-highest performance.
medical opticslandmark detectionconvolutional neural networksdeformable convolutionscoliosis
NETL
NSTL
万方数据
