ConvNet-Based Localization of Anatomical Structures in 3D Medical Images

摘要

Localization of anatomical structures is a prerequisite for many tasks inmedical image analysis. We propose a method for automatic localization of oneor more anatomical structures in 3D medical images through detection of theirpresence in 2D image slices using a convolutional neural network (ConvNet). A single ConvNet is trained to detect presence of the anatomical structure ofinterest in axial, coronal, and sagittal slices extracted from a 3D image. Toallow the ConvNet to analyze slices of different sizes, spatial pyramid poolingis applied. After detection, 3D bounding boxes are created by combining theoutput of the ConvNet in all slices. In the experiments 200 chest CT, 100 cardiac CT angiography (CTA), and 100abdomen CT scans were used. The heart, ascending aorta, aortic arch, anddescending aorta were localized in chest CT scans, the left cardiac ventriclein cardiac CTA scans, and the liver in abdomen CT scans. Localization wasevaluated using the distances between automatically and manually definedreference bounding box centroids and walls. The best results were achieved in localization of structures with clearlydefined boundaries (e.g. aortic arch) and the worst when the structure boundarywas not clearly visible (e.g. liver). The method was more robust and accuratein localization multiple structures.