CT to software and other considerations

Introduction

To design implants that are personalized to the patient’s anatomy and a specific surgical plan, engineers rely on medical imaging to create accurate anatomical reconstructions. CT image data is commonly required in orthopedic applications due to the volumetric dataset presented and clear contrast of bone, but MRI and X-ray data can also be used in unique cases. There are numerous commercially available software packages that use a similar process to convert a series of 2D medical images into 3D models. However, the focus of this chapter will be specifically on CT scanning (and its most common file format, DICOM), as it is commonly the ideal modality for most orthopedic applications.

Software workflow

Segmentation

To better explain the relevant CT scan parameters, it is useful to understand the software steps used to convert medical imaging into a 3D model. Once imported into an image viewer, DICOMs are segmented to model each bone as a separate entity. Images are viewed slice by slice and the contours of each bone are isolated through a combination of manual and automated processes. Hounsfield units, which describe radiodensity, are converted into grayscale values that allow for simple image processing. Thresholds are applied to the entire image stack to filter only the Hounsfield units of interest. Metal, bone, and soft tissue can quickly be separated from one another provided that the contrast remains uniform throughout the image. Any pixel in the scan volume that falls within the selected threshold is included and added to a binary mask. When all the pixels in a given mask are viewed in 3D space, the rough shape of the segmented anatomy can be visualized ( Fig. 2.1 ). Once thresholding is complete and bone has been isolated, it is still critical to separate the different regions of bony anatomy from one another. While some implants could be designed without the manual segmentation of each individual bone, it is often essential to visualize anatomical corrections.

Separating anatomy with similar Hounsfield units (such as separating different bones) is typically the most complicated part of segmentation. When joint spacing is ideal, region-growing algorithms can be used to separate one bone from another. Region-growing algorithms separate regions with similar contrast by “growing” out from one pixel and moving to a surrounding pixel if there is any connectivity to another pixel in the mask. Accordingly, if two bones are in adjacent pixels at any point of the joint, a region-growing algorithm will connect the two bones. When region growing is not possible, the contours of the bones are manually traced to separate bony regions. Depending on the quality of the CT scan, this tracing must be done on each individual slice ( Fig. 2.2 ).

Converting 2D to 3D

Once each bone has a corresponding mask that contains all the pixels of anatomy, it is converted into a 3D model using the pixel size and slice thickness of the CT scan. It is important to note that because bone masks are binary, there is no way to count partial pixels. To create the 3D models, the pixels in each bone mask are converted into a 3D cloud which is then used to generate a uniform, connected surface mesh for manipulation within a 3D modeling software package. This happens in multiple steps. Due to the nature of contrast in CT scans, the raw 3D point cloud from the bone mask will not be uniform, meaning that it will typically have holes, extraneous shells, and an inconsistent surface that does not match the bone. To convert this point cloud into a usable model, pixels that are not connected to the main shell are removed and the anatomy is often “wrapped” with a mesh that closes all the gaps in the model ( Fig. 2.3 ). The result is a uniform single surface that can be manipulated in a computer-aided design (CAD) software package. Surprisingly, the mesh surface that results from segmentation is surprisingly consistent, even between different software packages. When analyzing nine different commercially available segmentation software packages, the mean error in the triangle shape between the final meshes was 0.11 mm.

It is essential to convert the anatomy into meshes for several reasons. First, in medical applications 3D printers use mesh or triangle-based file formats (e.g., STLs) as their input files. To print anatomical models, all anatomy must be converted into meshes. Second, because patient-specific implants are designed directly from the surrounding anatomy, the models must be able to be manipulated freely in the design suite. In the case of a total talus, the starting point for design is the surface mesh of the contralateral talus. Once all relevant anatomy has been converted into a mesh that accurately represents the anatomy, implant design can begin ( Fig. 2.4 ).

CT scan considerations

Because binary masks from segmentation are directly used in design, there are several important considerations for CT images that make them suitable for preoperative planning and implant design. Both the specific scan parameters and scanning instructions used make a difference in the final quality of the devices designed.

Scanning parameters

Recommended CT scan parameters are provided by the implant manufacturer before 3D printing ( Fig. 2.5 ). In general, all recommendations are designed to minimize the size of each voxel, the smallest 3D volume element in a CT scan. In particular, it is ideal to minimize voxel size relative to the surrounding anatomy and maximize contrast per voxel.

Pixel spacing and slice thickness

Pixel spacing and slice thickness both define the smallest 3D volume or voxel size in a CT scan. Both parameters are defined in the scan protocol and can be readily extracted in the DICOM tags. Pixel spacing refers to the width of each pixel in the XY plane. Because clinical CT scans are taken axially, the pixel spacing always corresponds to the width of the pixels in the axial plane. Typically, this number does not need to be altered and is under 0.4 mm for most CT protocols. Slice thickness, or the distance between axial CT slices, is the second parameter that contributes to the smallest volumetric element in a scan and varies by CT protocol. Slice thickness typically ranges from 0.37 mm to 2.5 mm depending on the protocol used. Importantly, because this quantity is not the same as pixel spacing, CT images are rarely isometric, meaning that axial pixilation is visible in the initial surface models that are created from the CT scan ( Fig. 2.6 ). The difference in the reconstructed anatomy for a large-thickness scan is noticeable and impacts the final design. In some cases, a high slice thickness can make it almost the engineer’s best guess to model the outer contours of the implant ( Fig. 2.7 ).