A1147. AUTOMATIC SEGMENTATION OF OSTEOARTHRITIC KNEE JOINTS IN CT VOLUMES USING STATISTICAL BONE ATLASES

Abstract

Accurate segmentation of bone structures is an important step in surgical planning. Patient specific 3D bone models can be reconstructed using statistical atlases with submillimeter accuracy. By iteratively projecting noisy models onto the bone atlas, we can utilize the statistical variation present in the atlas to accurately segment patient specific distal femur and proximal tibia models from the CT data.

Our statistical atlas for the knee consists of 199 male distal femur models and 71 male proximal tibia models. We performed an initial registration between the average model from the atlas and the volume space before beginning the segmentation algorithm. Intensity profiles were linearly interpolated along the direction normal to the surface of the current model. The profiles were then smoothed via a low-pass filter. A point-tonearest peak gradient was calculated for each profile, and then weighted by a Gaussian window centered about the originating vertex. The flesh-to-bone edge locations are taken as the maximum of the weighted gradient. The detected locations were then projected onto the atlas using a subset of the available principal components (PC’s). The amount of variation is increased by projecting the edge locations onto a larger subset of PC’s. The process is repeated until 99.5% of the statistical variation is represented by the PC’s. Though our dataset is much larger, we initially performed bone segmentation on 5 male knee joints. The knee joint was considered to be the distal femur and proximal tibia. We used manually segmented models to determine ground truth. Initial results on the 5 knee joints (distal femur and proximal tibia) had a mean RMS error of 1.192 mm, with a minimum of 1.010 mm. Segmentation on the distal femur achieved a mean RMS error of 1.213 mm, and the results for the tibia had a mean RMS error of 1.264 mm.

Our results suggest that our atlas-based segmentation is capable of producing patient-specific 3D models with high accuracy, though patient-specific degeneration was often not well represented. To achieve more accurate patient-specific models, we must incorporate local deformations into the final model.