MORPHORMICS TECHNOLOGY OVERVIEW
Morphormics™ technology provides near-automatic segmentation of clinically important organs from CT images (and other imaging modalities).

The Problem
Solution Approach
Segmentation Algorithm Overview
Models
Initialization
Optimization Using the Image
Training
Summary

Morphormics™ technology provides near-automatic segmentation of clinically important organs from CT images (and other imaging modalities). The first application of this technology provides segmentation of organs of the male pelvis with application to radiation treatment planning for prostate cancer. This paper describes what the technology can do, how it works, and why this is an advance over previous approaches to segmentation.

THE PROBLEM
Three-dimensional conformal radiotherapy and intensity-modulated radiation therapy are preferred methods for treatment of prostate cancer. These treatment modalities require an accurate understanding of the three-dimensional anatomic configuration of the prostate and adjacent structures, including the bladder, rectum, seminal vesicles, and femoral heads. Accurate knowledge of the anatomic configuration, extracted from CT and other images, are critical inputs to three-dimensional treatment planning, improving the tumor-to-normal tissue dosage ratio. As a result, Morphormics technology improves treatment efficacy while simultaneously reducing treatment side effects.

The process of extracting the anatomic configuration from images is called segmentation. Conventionally, segmentation is accomplished by a physician or other highly skilled professional examining many (frequently 40 or more) individual image slices and manually drawing two-dimensional contours of each relevant organ in each slice. These 2-D contours are then stitched together to produce three-dimensional representations of the relevant organs, which are used as input to treatment planning.

Such contour-based manual segmentation consumes the (expensive) time of skilled and experienced clinical experts and yet produces results that vary considerably across the experts. Morphormics' near-automatic segmentation solution yields more consistent, accurate segmentations in much less time.

Figure 1. Selection of four initial points (X) on the boundary of the prostate (axial view).

SOLUTION APPROACH
Segmentation with Morphormics technology is a three-step process.

Step 1: Initial Point Identification
For each relevant organ, a small number of initial points on the surface of the organ are identified manually in the image. This step could be automated. We find, however, that fully acceptable clinical results are produced consistently when a small amount of human guidance is initially provided. Such initial guidance largely eliminates the manual post-processing that is often needed with competing methods. When segmenting the prostate, for example, four initial points on each of three axial slices are identified, as illustrated in Figure 1; for the bladder, a single interior point at the approximate center must be identified. We anticipate that typically a dosimetrist can identify these points.

Figure 1: Selection of four initial points (X) on boundary of the prostate.

Step 2: Automatic Segmentation
Given the image(s) and the initial points, an automatic algorithm performs the segmentation. This process takes from under a minute to approximately two minutes per organ, depending on organ type, image characteristics, and the initial points chosen. The algorithm computes the three-dimensional region comprising the organ (here, the prostate), as rendered in Figure 2. Of course, conventional two-dimensional contours can also be shown from the same computation (Figure 3) as the intersection of image slices and the three-dimensional volume.

Figure 2. Rendering of the volume identified by the algorithm as the prostate.

Figure 3. Two-dimensional contours through the segmentation produced by the segmentation algorithm. From left-to-right, panes show axial, sagittal, and coronal slices.

Step 3: Evaluation and Editing
The user examines the resulting segmentation, deciding whether to accept or to edit the segmentation produced. In standard clinical practice the radiation oncologist must approve the final results before they are used for planning. We expect that editing will only rarely be necessary, perhaps in 10% of cases. When editing is required, the user will be able to choose from multiple methods: Frequently, adding more initial points and re-running the segmentation algorithm will do the job. Conventional contour editing is also available.

Importantly, the result of this three-step segmentation process is an accurate three-dimensional representation of the relevant organs and their relative positioning in space, not a set of disconnected two-dimensional contours. Both the shape information and spatial relationships are illustrated in the labeled rendering (from a different patient) shown in Figure 4.

Figure 4. Rendering of three-dimensional organ shapes and spatial positioning.

SEGMENTATION ALGORITHM OVERVIEW
To introduce the algorithm's model-based approach, consider the knowledge that an expert human brings to bear when doing a manual segmentation. The human expert understands the organs of interest, the range of shapes each may assume under various conditions (e.g., bladder empty or full, distension of the rectum), the range of spatial relationships among the organs, the typical image intensities associated with the organs and other material (surrounding tissue, fat, urine, fecal matter, and gas) and various visible landmarks in the images. Morphormics' algorithms model this information, without which automatic methods are unlikely to be adequate compared to an expert.

MODELS
The algorithm is classified as model-based because it uses models that capture the kinds of information used by human experts. In particular, we use two kinds of models:

The shape model is defined using m-reps ("m" for "medial"), a shape representation of an object as a set of medial atoms, each of which consists of two spokes emanating from a hub in the "middle" of the object (hence, the name "medial"), as illustrated in Figure 5. M-reps have many interesting and useful mathematical properties. They are particularly suitable as a shape model for segmentation because they can capture both local variations in shape as well as more global deformations. The numbers that characterize a set of hub locations and spoke lengths and orientations completely define a shape and thus can be used as a parameterization of a particular shape. To reduce the number of parameters needed to capture the shape space for an organ, a mathematical technique is applied to find the subspace in which the most significant variation takes place, i.e., the shape space. The key point is that using m-reps, any credible organ shape can be defined within the shape space by a small set of parameters, which improves processing efficiency, and solutions are largely limited to only credible variants of the organ.

Figure 5: A medial atom (left) and the implied boundary of the object, a bladder (right), defined by the m-rep.

The image intensity model constructs, using the shape model, a set of local object-interior and object-exterior regions for each organ. For example, in Figure 6, a variety of regions are defined relative to the boundary of a prostate. For each of these regions, the image intensity model uses a representation of the region's intensity histogram to provide the algorithm with information about the variation in intensity that is expected in the region.

Figure 6: Image regions defined relative to the boundary of a prostate shape model.

INITIALIZATION
The algorithm begins with the initialization phase, in which the appropriate organ shape model is coarsely positioned, oriented, and deformed. This is based either on initial points given by the user, initial points obtained by automatic detection of high contrast points on the organ surface, rigid motions of m-rep models resulting from segmentations of previous images of the same patient, or by reference to previous initialization (or segmentation) of an adjacent organ in the same image set.

OPTIMIZATION USING THE IMAGE
Once the shape model is appropriately positioned, the regions defined by the image intensity model can be matched to the target image and the corresponding intensity data gathered. An objective function is defined based on the shape and image intensity models that measures the probability that both the shape model corresponds with what is known about the distribution of shapes in the shape space and that the intensity patterns in the image are consistent with the anatomy and how it varies. To improve the match and converge to a final segmentation, a multi-dimensional optimization algorithm is run to adjust the model parameters to maximize the objective function, i.e., to find the most probable variant of the anatomy, given the image.

TRAINING
As described above, an important role of the shape and image intensity models is to capture information that a human expert would apply when doing manual segmentation. This requires a training process to obtain information about the variations that are expected across patients, imaging devices, and within patients over time. In training, several experts manually segment the same relevant organs in the same sets of raw image sets, providing a robust sample size of patients' organs. This information is then processed to extract the probability information that characterizes the shape space for each organ and the corresponding image intensity distributions. This information is stored ahead of time and is later used in producing segmentations.

SUMMARY
Morphormics technology provides high-quality, rapid, near-automatic segmentation of organs in CT and other images. Applications include radiation treatment planning and other clinical and research uses. These capabilities will be provided in FDA-cleared medical devices, with associated expert training image datasets. Our use of trainable models makes the algorithm particularly resilient to variations in image quality and to across-patient variability. In rare instances when the clinician judges that the computed segmentation needs improvement, it is easy to provide additional guidance to the algorithm as a way to quickly "edit" the segmentation result.

Morphormics technology is not presently for sale in the United States.