Show abstract - PDF - Reviews

Accurate detection of 3D landmarks is critical for evaluating and characterizing anatomical features and performing preoperative diagnostic screening. However, detecting 3D landmarks can be challenging due to the local structural homogeneity of medical images. To address this issue, physicians often annotate multiple landmarks in a single slice, particularly when estimating 3D distance or volume. In this study, we present a prior guided coarse-to-fine framework for efficient and accurate 3D medical landmark detection; we make use of the prior information that in specific settings physicians annotate multiple landmarks on a same slice. The coarse stage uses coordinate regression on downsampled 3D images to maintain the structural relationships across different landmarks. The fine stage categorizes landmarks as correlated and independent landmarks based on their annotation prior. For independent landmarks, we train multiple models to capture local features and ensure reliable local predictions. For correlated landmarks, we mimic the manual annotation process and propose a correlated landmark detection model that merges information from various patches to query key slices and identify correlated landmarks. Our method is extensively evaluated on two datasets, exhibiting superior performance with an average detection error of respective 3.29 mm and 2.13 mm.
Hide abstract