Download Automated Microct-Based Bone And Articular Cartilage Analysis Using Iterative Shape Averaging And Atlas-Based Registration Book in PDF, Epub and Kindle
INTRODUCTION: Micro-computed tomography (u00b5CT) is a gold standard modality for 3D assessment of bone. Contrast-enhanced (CE)-u00b5CT is emerging as a powerful technology for 3D characterization of articular cartilage (AC) tissue, historically assessed via 2D histology. Volume-of-interest (VOI) selection of these tissues generally involves manual contouring, which is time consuming and leads to inter- and intra-observer variability. AC is also often analyzed in tandem with adjacent subchondral bone, further compounding these issues. The limitations of manual contouring can be mitigated by the application of semi-automated or fully automated segmentation, but given the inherent complexity of the joint tissues being imaged, few automated segmentation techniques have been described for u00b5CT imaging. Atlas-based tissue segmentation is an effective technique in which a pre-segmented, representative set of tissue volumes (an u201catlasu201d) is generated and registered onto an unknown image, automatically segmenting the tissues of the unknown image. Here we demonstrate the use of iterative shape averaging (ISA) to generate a multi-tissue atlas to automatically segment and analyze multiple tissues from u00b5CT data.METHODS: Under IACUC-approved protocols for unrelated studies, dissected distal femora from 50 adult Lewis rats (25 healthy limbs, 25 post-ACL rupture) were incubated in the contrast agent ioxaglate and underwent CE-u00b5CT imaging (u00b5CT-40, Scanco Medical) with 12 u00b5m voxel size. An experienced user performed manual VOI selection of epiphyseal trabecular bone and AC using manual contouring followed by intensity thresholding. These samples were then randomized into training sets of 1, 3, 5, 10, and 20 samples (n=10 sets/group), and all other samples in the set were assigned to the test group. For the 1-sample atlas, a randomly selected sample was chosen as a representative atlas and registered onto test images. For all other groups, the training images were used to generate an average atlas using ISA (Fig. 1A). Briefly, training images were rigidly co-registered onto a random reference image, and an initial average was created. Subsequent rounds of non-rigid registration (NRR) iteratively refined the average until it converged on an average image and corresponding average multi-tissue atlas. Each atlas was then registered onto 10 randomly selected test images to segment epiphyseal bone and AC volumes. Segmentations were assessed against manual VOIs in terms of Dice similarity coefficient (DSC), sensitivity, and specificity. Bone densitometry and morphometry and AC morphology measures were also calculated from each volume, and percent (%) error was calculated between atlas-based and manual segmentations. Differences were assessed via one-way ANOVA with a Bonferroni correction for multiple comparisons.RESULTS: Mean DSC and sensitivity for epiphyseal bone and AC increased with the number of input samples to the atlases, particularly DSC of AC tissue (Fig. 1B). Percent error for bone parameters and AC volume was within u00b15%, with decreased variance in atlases built from 10 or 20 training samples. Percent error in AC thickness was higher but remained within u00b110% (Fig. 1D). Atlases with 1, 3, and 5 inputs had rare cases of gross mal-registration of both bone and AC (Fig. 1C, D), whereas those from 10 and 20 inputs had no cases of mal-registration. The rate of highly accurate registration (DSC > 0.85) generally increased as more inputs were added (Fig. 1C).DISCUSSION: Automated segmentation of AC and bone from u00b5CT scans can be performed on healthy and degenerate femora with high accuracy via ISA and atlas-based segmentation. Atlases built from few samples were capable of accurate registration in most cases but were more prone to mal-registration. Incorporating more training images produced a more robust atlas, and 10 training images were sufficient to eliminate mal-registration. Bone morphometric and densitometric parameters all fell within u00b15% of those from manual segmentation. AC thickness was more variable u2013 within u00b110% error u2013 likely due to the much smaller tissue volume compared to epiphyseal bone. AC is generally adjacent to thicker tissues including ligament insertion sites and other soft tissues, and small differences in the tissue border can significantly impact resultant volume and thickness. Many parameters consistently exhibited positive or negative error, suggesting bias between manual and atlas-based segmentation, rather than inaccuracy, as the main source of error. In AC, where thickness increases near soft tissue borders, negative error in thickness and volume suggests that borders on atlas-derived VOIs may be more conservative than manual VOIs. This should not be taken to represent a weakness in either atlas-based or manual VOI selection, but rather an inherent difference in the two strategies.SIGNIFICANCE: Tissue segmentation is a highly time- and labor-intensive process and a source of variability in both preclinical and clinical imaging studies of joint tissues. Our work demonstrates that atlas-based segmentation could be a powerful tool to enable significantly faster and more consistent joint tissue segmentation of both healthy and degenerate limbs, accelerating the rate and sensitivity of imaging-based research.