We present a fully automatic method of segmenting and landmarking hip CT images for planning of Total Hip Arthroplasty (THA). Our method consists of two stages, i.e., the segmentation stage and the landmarking stage. At the segmentation stage, a multi-atlas segmentation constrained graph method is employed to fully automatically segment both the pelvis and the bilateral proximal femurs from the input CT data. The segmentation stage is followed by the landmarking stage, where a set of pre-defined landmarks are transferred from generic models of the associated hip structures to the input CT space via non-rigid registrations in order to compute a set of functional parameters that are relevant to planning of THA. Evaluated on 20 hip patients, we computed both the segmentation accuracy and the landmarking accuracy. An average segmentation error of 0.38 ± 0.25 mm and 0.49 ± 0.22 mm was found for the hemi-pelvis and for the proximal femurs, respectively. For 3D landmarking, a mean error of 1.58 ± 0.87 mm and 0.46 ± 0.39 mm was found for the acetabular rim center and the acetabular rim radius, respectively; a mean error of 0.74±0.45o was found for the orientation of the anterior pelvic plane; and a mean error of 3.14 ± 1.90 mm and 2.04 ± 1.61 mm was found for the femoral head center and the femoral offset, respectively.
Segmentation of Lung Ct Images using Cascaded Fully Convolutional Neural Networks