Region-growing algorithm

Abstract

A technique for automatically generating a virtual model of a branched structure using as an input a plurality of images taken of the branched structure. The technique employs an algorithm that avoids inaccuracies associated with sub-optimal threshold settings by “patching” holes or leaks created due to the inherent inconsistencies with imaging technology. By “patching” the holes, the algorithm may continue to run using a more sensitive threshold value than was previously possible.

Description

BACKGROUND OF THE INVENTION

Navigating the airways of the lungs has always presented challenges to physicians attempting to diagnose and treat lesions transluminally. As such, numerous navigational aids and imaging tools have been developed and/or utilized to provide physicians a “map” of the lungs.

One such tool is a CT scanner. CT scanners use X-ray technology to take multiple scans or “slices” of the lungs. These scans each represent a cross-section of the lungs and can be viewed individually or assembled, via computer programs, to form a three-dimensional CT model. However, CT scans, like most images using X-ray technology, are somewhat cloudy and translucent in nature and difficult to view. As such, computer graphics techniques are employed to interpret the information provided by the CT model and “grow” a virtual model of the airways which mimics what might be seen by a bronchoscope navigating the airways. An example of this process is shown and described in U.S. patent application Ser. No. 11/939,537, entitled Adaptive Navigation Technique For Navigating A Catheter Through A Body Channel Or Cavity, the entirety of which is incorporated by reference herein.

This graphical technique is sometimes referred to as “region growing or 3D map generation,” and presents its own challenges. For example, region growing typically involves a processing of the CT data by analyzing each two-dimensional pixel, or, more pertinently, three-dimensional voxel, for brightness or “density” and assigning the voxel a value that indicates either tissue or air based on whether the density meets a certain threshold value. CT scans are grayscale images composed of a plurality of pixels (2D) or voxels (3D—if the scans have been assembled into a volume), each pixel or voxel varying in intensity from white (most intense) to black (least intense). Each intensity level between white and black appears as a shade of gray. By designating the various shades of gray from the CT scans either “tissue” or “air” the resulting image of the lungs becomes much more clear. However, if the voxels are designated incorrectly, the model of the lungs becomes inaccurate. Incorrect voxel designation results from setting the threshold level at an incorrect value, which is an inherent problem when attempting to assign discreet values (air or tissue) to voxels which are actually various shades of gray.

A presently-used technique for optimal threshold setting is beginning with a conservative threshold and performing a region-growing iteration. A conservative threshold is one that is not likely to result in leakage, which occurs when tissue is designated as air and creates a virtual image of the airways that looks like air (color) is spilling out of the airways. However, even with a conservative threshold, inaccuracies in the CT scans can result in “holes” after the segmentation process. These holes result in false branches.

Moreover, a conservative threshold results in airways that end prematurely. Therefore, after a conservative iteration is performed, resulting in a stunted branched structure, the threshold is incrementally increased and a second iteration is performed. These steps are repeated until leakage occurs. Thus, the maximum threshold that does not result in leakage is determined and used. This approach, however, naturally results in the least-dense portion of the CT image dictating the threshold level. There are other problems that arise from this method as well.

For example, during a full inhalation, the airways stretch and thin in order to accommodate the additional air volume. The thinning of the airways results in reduced tissue imaging density, and leakage thus arises even at lower threshold values.

Another example is that each time the threshold is increased, the algorithm runs from the initial seed point. Hence, if the threshold is increased ten times before leakage arises, each of the voxels analyzed in the initial iteration, is analyzed nine more times. This algorithm is thus inherently taxing on processing resources.

As such, a need is identified for a region-growing algorithm that is able to identify localized weaknesses in image data and “repair” them such that more distal branches of a bronchial tree can be segmented and “grown”.

SUMMARY OF THE INVENTION

The present invention provides a technique and algorithm for controlling leakage while using a high threshold during a region growing procedure. Rather than beginning with a conservative (low) threshold, a high threshold is used but growing is analyzed and controlled.

One aspect of the present invention uses an algorithm that defines iteration as a single voxel layer of growth, rather than growth until leakage is detected. Each of the iterations is analyzed for a predicted increase (e.g., parabolic increase) in the number of voxels. If the voxels are not growing according to an expected rate, the unexpected increase is identified as leakage.

Rather than ending the region-growing process when leakage is detected, the point of leakage is isolated or “patched” so that further iterations on the current path do not continue from that point, but the rest of the model is allowed continued growth.

Another aspect of the present invention provides a post-iteration step which, in effect, repairs or salvages branches encountering leakage. After all of the iterations are performed, branches that had leakage are retrieved from a stored location and regrown with a reduced threshold. These branches are then added to the model.

The advantages of the algorithm of the present invention are numerous. One advantage is a more meaningful increase in the number of branches as a result of the utilization of the highest possible threshold.

Another advantage of the present invention is that the appearance of “holes” in the central segmented area of the model is greatly reduced.

Yet another advantage of the present invention is a reduced number of false branches. Because each iteration is analyzed, growth is controlled and false branches are prevented from growing.

Another advantage of the present invention is that the branches grown have an increased length, due to an optimized threshold as well as secondary growing efforts.

Still another advantage of the present invention is reduced importance on initial threshold value selection. Through experimentation, it has been found that this algorithm is self-correcting, yielding virtually identical results regardless of the initial threshold used, so long as the thresholds are relatively high.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow-chart of a preferred embodiment of an algorithm of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

The technique 10 of the present invention is charted in the flowchart presented as FIG. 1. The technique 10 begins at 20 by selecting a starting point for the segmentation of the CT data. For example, selecting a point in the trachea is a logical starting point as it is the largest, and most proximal airway, and easily recognizable on a CT scan. Preferably, the starting point is a single voxel inside the trachea that is centered and meets a high threshold value. In other words, a voxel which is clearly air is selected. Any point from which further adjacent voxels are analyzed is hereinafter referred to as a “seed point”. The starting point is the first seed point.

At 30 the propagation process is initiated by designating adjacent voxels around the starting point. This designation is known as segmentation, and it indicates that the new voxels met the threshold level for air. Because the starting point is preferably selected in the trachea, and it is not desired to grow the airway back toward the mouth of the patient, only voxels in a distal direction of the starting point are segmented. As such, a “wave” is generated that travels distally into the airways. It is understood that the branches of the lungs fan out in all directions. As such, “distal direction” as used herein is interpreted as getting further away from the starting point along a path that remains in the airway. In other words, some “distal” points in the airways may be relatively close to the starting point if one were to cut through tissue to make a straight line between the points.

It is understood that the starting point, being a single voxel, is surrounded by 26 adjacent voxels, 17 of which are extending in a direction of desired growth (in a direction not yet segmented and not in a reverse direction, such as toward the mouth).

At 40, it is determined whether any new voxels were segmented. In other words, if no voxels in that “wave” met the threshold level for air, there is no growth and the tree is complete.

If the answer at 40 is “yes”, the algorithm continues to 50, where the segmented voxels are counted and analyzed to determine whether there is leakage. The new voxels may be more numerous than the previous iteration, but the growth should be controlled. In other words, if the increase in the number of voxels is expected (more specifically, a parabolic increase has been observed in normal growth patterns without leakage) then it is determined that there is no leakage and the process returns to step 30 to begin a new iteration. For example, as indicated above, the starting point consisted of one voxel. Assuming the starting point was surrounded by air, the next “wave” of voxels would be seventeen new seed points. However, because many of these are adjacent to each other, the next successive wave would not give rise to seventeen new seed points for each of the previous seventeen seed points. Additionally, the voxels behind each seed point that have been already analyzed, are not segmented again. If the airway being segmented were perfectly cylindrical, as soon as the seed points reached the walls of the airway, the “wave-front” would be a convex sheet, a single voxel in thickness, and would remain constant. Hence, the mathematical model for growing is somewhat parabolic, except when new branches are introduced, and considering that the airways are narrowing in the distal direction. Leakage, however, results in an abrupt increase in the number of segmented voxels.

If at 50 the analysis results in an unexpected or abnormally high increase in segmented voxels, it is determined that leakage exists and the process moves to step 60, which identifies and records the segmented voxels from the previous iteration and labels them as accurate.

Leakage determination is derived from two important conclusions: (1) It is expected that the front size has an inverse (not necessarily linear) dependence on the iteration number, e.g. [front size]˜f(1/iteration number). (2) Bifurcations and changes in airway shape may result in somewhat linear growth in front size.

At 70, upon the detection of leakage, an analysis is done in order to select the most recent “good” iteration that does not contain leakage. This decision is based on the same principals used to satisfy the mathematical model in compliance with the natural structure of the bronchial tree.

The voxels segmented up to the last good iteration are assumed to be committed to the segmented voxel list, while voxels that belong to each iteration above the “good” one are analyzed in order to separate the voxels that led to the leakage. In order to make this analysis, recently segmented voxels are labeled to connected objects (part of the branch). Each object is then analyzed in order to detect the object that caused leakage.

The coordinates of voxels that belong to inaccurate objects are stored separately and treated differently thereafter. They are locked in the binary map to prevent their participation in the segmentation process. Voxels, belonging to accurate branches are returned to the segmentation and the process returns to 30.

At 80 the objects identified as leaking are removed from further segmentation and stored in a queue for post processing. After 80 the process returns to 30 for the next iteration.

If at 50 the answer was “no” for leakage, the process returns to 30 for the next iteration. It should be noted that the flow chart 10, though presented in series for clarification, is actually a parallel operation. In other words, each new voxel is a seed point and the flow chart is performed on each next iteration therefrom simultaneously. Hence, viewing the growth of the bronchial tree real time, one would see a near-instant tree appear, depending of course on the power of the processor running the algorithm.

If at 40 there are no new voxels detected, the algorithm proceeds to 90, which is a decision step asking whether any leakage objects were identified. If the answer is “yes” step 100 is executed, which retrieves the last object from the storage queue (step 80).

Next at 110, the threshold is reduced and the algorithm is performed on only the selected leakage object. Because a reduced threshold is being used, the likelihood of leakage is reduced.

If at 90 it is determined that there are no leakage objects, either because there were none or they have all been reprocessed, the process is completed at 120.

Although the invention has been described in terms of particular embodiments and applications, one of ordinary skill in the art, in light of this teaching, can generate additional embodiments and modifications without departing from the spirit of or exceeding the scope of the claimed invention. Accordingly, it is to be understood that the drawings and descriptions herein are proffered by way of example to facilitate comprehension of the invention and should not be construed to limit the scope thereof.

Claims

1. A region growing method comprising: identifying, by a processor coupled to a memory, a voxel as a first seed voxel in a wave;performing, by the processor, region growing iterations with the first seed voxel in the wave until there are no newly segmented voxels in a successive wave and no leakage voxels left in a post processing queue, each region growing iteration including:identifying, by the processor, voxels adjacent to a seed voxel in a current wave as seed voxels in a successive wave, when each of the adjacent voxels meets or exceeds a threshold value;determining, by the processor, whether a seed voxel in the current wave is a leakage voxel;when the seed voxel in the current wave is determined to be the leakage voxel,removing, by the processor, the seed voxel from the current wave and storing the seed voxel as a leakage voxel in the post processing queue;determining, by the processor, whether the successive wave includes at least one seed voxel;when the successive wave is determined to include no seed voxel, determining, by the processor, whether the post processing queue includes at least one leakage voxel; andwhen the post processing queue is determined to include at least one leakage voxel, selecting, by the processor, a leakage voxel from the post processing queue as a seed voxel in the successive wave and reducing the threshold value so that the reduced threshold value is used as a new threshold value in next region growing iterations.

2. The region growing method of claim 1, wherein the first seed voxel is a point in a trachea of a patient.

3. The region growing method of claim 2, wherein the first seed voxel in the trachea meets a high threshold value.

4. The region growing method of claim 1, further comprising counting a number of seed voxels in the successive wave.

5. The region growing method of claim 4, wherein the seed voxel in the current wave is determined to be a leakage voxel when an increase in the number of seed voxels in the successive wave exceeds a predefined rate.

6. The region growing method of claim 5, wherein the predefined rate is related to an inversely proportional relationship of a front size of adjacent voxels and an iteration number and (b) a linear relationship between the front size and bifurcations and changes in an airway shape.

7. The region growing method of claim 1, wherein, when no seed voxels in the current wave are determined to be a leakage voxel, the region growing method proceeds to a next region growing iteration.

8. The region growing method of claim 1, further comprising generating a model from seed voxels in the current and successive waves.

9. The region growing method of claim 8, wherein the model is of an anatomical structure of a patient.

10. The region growing method of claim 9, wherein the anatomical structure is an airway starting from a trachea to a lung.

11. The region growing method of claim 1, wherein an initial threshold is a Hounsfield unit value for air.

12. The region growing method of claim 1, wherein the method is performed by a processor.