Title: 3D articular cartilage reconstruction using in vivo multinuclear mr images
Abstract: Purpose: Morphology (thickness and volume) of articular cartilage (AC) plays an important role in assessing onset and progression of osteoarthritis (OA). Thickness and volume of AC can be measured in different compartments of weight-bearing and central regions that require reconstructions of 3D AC models. However, due to local variations in AC tissue properties, knee MR images obtained from most hydrogen based MR pulse sequences suffer from low and varying intensity issues that makes it difficult to segment AC automatically for subsequent 3D reconstruction. This work aims to reconstruct 3D models of AC by the combined assessment of 23Na and 1H multinuclear in vivo knee MR images. Methods: A specialised radiofrequency (RF) coil consisting of two channels (23Na & 1H) is used to scan four human subjects at 1.5T MR system that facilitates MR spectroscopy imaging. Two separate MR pulse sequences namely, 3D Gradient Echo sequence and MEDIC are used for sodium and proton imaging of knee respectively. Data processing has been performed using customised routines written in MATLAB®. Pre-processing of sodium MR images is performed to extract sodium rich region by re-sampling of sodium extracted slices to interpolate sodium slices located in the field of view corresponding to proton slices. The product and average of re-sampled sodium extracted slices with original proton slices produces fused MR images. Automatic segmentation of AC is then performed in which enhanced cartilage regions in the fused images are segmented using thresholding (Otsu algorithm). Here, all connected components (set of pixels that form a connected group) in the binary threshold image are labelled resulting in the identification of all connected groups as separate objects in the labelled image. Since cartilage is typically convex in shape, the largest convex shaped object in the image is extracted in the following step. The edges of the extracted object form the segmented AC region. The flow diagram of Fig. 2 represents the steps involved in automatic segmentation. For the reconstruction of 3D AC models, all segmented slices are stacked in parallel and isosurfaces are computed to extract vertices and the faces that connect points of equal elevation. The data in rendered in 3D space to visualise the 3D AC models. The flow diagram for the steps involved in 3D reconstruction is also shown in Fig. 2. Results: Figure 1 shows the visual representation of results obtained at different stages involved during pre-processing. Figure 1(a) and Figure 1(b) represents the original proton and sodium knee MR images respectively that are acquired using a specialised RF coil. Sodium knee MR images are then further processed and sodium rich region is extracted as shown in Fig. 1(c). Since, the size of sodium extracted slices and original proton slices are different, up-sampling using cubic spline interpolation is applied on sodium extracted slices resulting in similar size to proton MR images as shown in Fig. 1(d). Furthermore, fusion of re-sampled sodium extracted slices with original proton slices is performed by localising slices in field of view and taking the product and average of corresponding slices as shown in Fig. 1(e). The resultant fused slice contains high contrast in AC region compared to surrounding tissues that were further segmented automatically by applying threshold and image property operations as shown in Fig. 2 (a). In this way, all the slices are first segmented automatically and further stacked in parallel. Segmented slices are further processed to extract the vertices and faces using isosurface and rendering is performed to generate 3D models of AC. In addition, smoothening is applied on vertices and faces for better rendering and visual representation of AC 3D models as shown in Fig. 2(b). Fig. 2Left: Flow diagrams showing steps involved in automatic segmentation and corresponding illustration showing the outputs (a). Right: (a) Flow diagram showing step involved in 3D reconstruction and corresponding reconstructed models (top) and after applying smoothening (bottom).View Large Image Figure ViewerDownload Hi-res image Download (PPT) Conclusions: The use of multinuclear (23Na and 1H) MR knee data can overcome the low and varying intensity in MR images. This approach enables AC regions to be automatically segmented with advanced image processing techniques. Reconstructed 3D articular cartilage models can be used to assess morphology of AC in different compartments.