Processing of musculoskeletal dynamic computer tomography images: a multi-atlas segmentation approach

Purpose: We propose an automated workflow for segmentation of bony structures in dynamic musculoskeletal CT images basedon a multi-atlas segmentation (MAS) approach. We evaluated ourworkflow on dynamic sequences of the thumb base and the kneejoint.Materials and Method: 15 healthy volunteers were subjected todynamic CT scanning (256 slice Revolution CT, GE Healthcare)while performing instructed cyclic joint movements: oppositionrepositionmovement of the thumb (5 volunteers) and flexionextensionof the knee (10 volunteers). Static scans of the joint inneutral positions were also acquired to create the atlas dataset (imageand its corresponding segmentation/label). In each dynamic dataset,an image with the joint in neutral position was selected as referenceimage. The bones in this reference image were segmented using theMAS approach. Atlas labels were created by manual segmentationof the bones in collaboration with an expert in bone anatomy usingITK Snaps’s active contour mode. The patella, femur and tibia were segmented for the knee images, for the thumb we considered firstmetacarpal bone and the trapezium. The MAS approach consistedof a three-step process: 1) a pairwise registration of the referenceimage to the set of atlases to find optimal transformations thatalign each atlas to the reference image; 2) the propagation of theatlas labels onto the reference image using the correspondingtransformation from step 1; 3) a label fusion step which fuses alllabels into a signal final segmentation. The MAS was validated bya leave-one-out cross-validation experiment on each joint dataset.Success of the segmentation was evaluated using overlap measures(dice coefficient, false positive (FP) and false negative (FN)) volumeerror fractions).Results: The proposed MAS approach resulted in Dice coefficientscores of 0.89 ± 0.02 for the thumb dataset and 0.94±0.05 for theknee. FP was 0.08 ± 0.04 and 0.05 ± 0.03 while FN was 0.14 ± 0.05and 0.06 ± 0.03 for the thumb and knee data respectively.(Figure Presented)Conclusions: We setup an automated workflow for segmentation of joints undergoing motion during dynamic CT scanning. Our work contributes to an automated workflow, and hence clinical feasibility, for a quantitative motion analysis of dynamic CT MSK data