Fully Automatic Segmentation of Lumbar Vertebrae from CT Images using Cascaded 3D Fully Convolutional Networks

摘要

We present a method to address the challenging problem of segmentation oflumbar vertebrae from CT images acquired with varying fields of view. Ourmethod is based on cascaded 3D Fully Convolutional Networks (FCNs) consistingof a localization FCN and a segmentation FCN. More specifically, in the firststep we train a regression 3D FCN (we call it "LocalizationNet") to find thebounding box of the lumbar region. After that, a 3D U-net like FCN (we call it"SegmentationNet") is then developed, which after training, can perform apixel-wise multi-class segmentation to map a cropped lumber region volumetricdata to its volume-wise labels. Evaluated on publicly available datasets, ourmethod achieved an average Dice coefficient of 95.77 $\pm$ 0.81% and an averagesymmetric surface distance of 0.37 $\pm$ 0.06 mm.