Difference between revisions of "Projects:NerveSegmentation"

From NAMIC Wiki
Jump to: navigation, search
 
(17 intermediate revisions by 3 users not shown)
Line 3: Line 3:
 
= Nerve Segmentation =  
 
= Nerve Segmentation =  
  
Automatic segmentation of neural tracts in the dural sac and outside of the spinal canal is important for diagnosis and surgical planning. The variability in intensity, contrast, shape and direction of nerves in high resolution MR images makes segmentation a challenging task. In this paper, we present an automatic tracking method for nerve segmentation based on particle filters. We develop a novel approach to particle representation and dynamics, based on Bezier splines. Moreover, we introduce a robust image likelihood model that enables delineation of nerve bundles and ganglia from the surrounding anatomical structures. We demonstrate accurate and fast nerve tracking when compared to expert manual segmentation.
+
Automatic segmentation of neural tracts in the dural sac and outside of the spinal canal is important for diagnosis and surgical planning. The variability in intensity, contrast, shape and direction of nerves in high resolution MR images makes segmentation a challenging task.  
  
 +
In this project, we present an automatic tracking method for nerve segmentation based on particle filters. We develop a novel approach to particle representation and dynamics, based on Bezier splines. Moreover, we introduce a robust image likelihood model that enables delineation of nerve bundles and ganglia from the surrounding anatomical structures. In the results section below, we summarize accurate and fast nerve tracking when compared to expert manual segmentation.
  
 
= Data =
 
= Data =
  
[[File:NerveSegMREg.png|800px|thumb|center|We show three slices of an example High-Resolution MR scan. All arrows point to the same nerve bundle. Blue arrows show examples of poor contrast between the nerve and the surrounding tissue; orange arrows indicate the thickening of the neural tract into a ganglion.
+
[[File:NerveSegMRE2.png|800px|thumb|center|We show four slices of an example High-Resolution MR scan. The blue outline in the first figure shows a manual segmentation of a nerve for reference. All arrows point to the same nerve bundle. Blue arrows show examples of poor contrast between the nerve and the surrounding tissue; orange arrows indicate the thickening of the neural tract into a ganglion.  
 
]]
 
]]
  
 
= Description =
 
= Description =
In this paper, we present a tracking approach based on particle filtering, also known as sequential Monte Carlo tracking. Tracking has also been used successfully for segmentation of tubular structures. Most vessel tracking methods model the state as a cross-sectional ellipse or as a cylindroid. In tracking nerve bundles, the regions of low contrast require the state to capture substantially longer segments of the track than what is represented by a cross-section. In addition, nerves tend to change direction, often sharply, which necessitates a use of more complex descriptors than cylinders.  
+
We present a tracking approach based on particle filtering, also known as sequential Monte Carlo tracking. Tracking has also been used successfully for segmentation of tubular structures. Most vessel tracking methods model the state as a cross-sectional ellipse or as a cylindroid. In tracking nerve bundles, the regions of low contrast require the state to capture substantially longer segments of the track than what is represented by a cross-section. In addition, nerves tend to change direction, often sharply, which necessitates a use of more complex descriptors than cylinders. Several of the data features can be seen in the Data section above.
  
To address the challenges of nerve tracking, we define a flexible particle representation that captures the geometric behavior of the nerve bundles. We use a Bezier spline centerline with a linear radius function to characterize a nerve bundle. We devise a dynamics model for particle updates that encourages continuity and smoothness. Furthermore, we define an image likelihood model that compares gradient fields and intensities of predicted patches with image observations to evaluate a posterior distribution of the particles' importance. Once tracking is completed, we remove spurious segmentations by measuring the quality of the entire tract. We demonstrate successful segmentations of neural tracts and evaluate them relative to expert manual segmentations. To the best of our knowledge, this is the first automatic segmentation of nerve bundles and ganglia.
+
To address the challenges of nerve tracking, we define a flexible particle representation that captures the geometric behavior of the nerve bundles. We use a Bezier spline centerline with a linear radius function to characterize a nerve bundle. We devise a dynamics model for particle updates that encourages continuity and smoothness. Furthermore, we define an image likelihood model that compares gradient fields and intensities of predicted patches with image observations to evaluate a posterior distribution of the particles' importance. With the particle model, dynamics model and likelihood model, we can implement a full particle filter. Once tracking is completed, we remove spurious segmentations by measuring the quality of the entire tract.  
 +
 
 +
To the best of our knowledge, this method is the first automatic segmentation of nerve bundles and ganglia.
  
 
= Results =
 
= Results =
  
[[File:NerveSegResults.png|800px|thumb|center|
 
The above figure shows a summary of our results. On the left, we show a rendering of segmentation results -- the rightmost nerve shows results without post-processing pruning, while the left segmentation was processed after completion of tracking with an automatic post-processing step. On the right, we present two measures, precision and sensitivity, as box plots.]]
 
  
Here we compare our segmentations with expert delineations. We define precision as the percentage of voxels identified by the algorithm as nerve that were also marked as nerve by an expert. Precision coarsely quantifies our ability to separate the nerve from the surroundings. We also compute precision for a dilation of the manual segmentation by two and three voxels. The results for precision are shown in the left part of the graph, where we note that we reach a median $97.3\%$ of our proposed segmentation being within 2 voxels of the expert segmentation, while due to partial volume effect we over-segment the thin nerves and we only achieve a 68% precision with respect to exact segmentation.
 
  
Since the main goal of nerve tracking is to extract the full path of the nerve accurately, while it is less crucial to identify the bundle boundaries, we also define sensitivity to quantify the agreement between the automatically defined path and the centerline of the manual segmentation. We define sensitivity as the percentage of voxels in the centerline of the expert segmentation that were also correctly identified by our algorithm. We also compute the same percentage for a dilation of the automatic segmentation, by two and three voxels. We achieve a media 96% of the centerline voxels being within 3 voxels of a prediction, while we find a drop to 53% when comparing with the exact centerline. The latter drop is due to segmentations in the areas of the thick ganglia, where our method under-segments, and may follow a slightly off-center path.
+
 
 +
[[File:NerveSegResultsGraphs.png|800px|thumb|center|
 +
Left: Summary of synthetic results for the eight types of synthetic nerves and backgrounds. The
 +
red bars indicate the average of the median distance between the automatic and the manual outline. The
 +
blue bars indicate the 90th percentile. For each category, Model indicates that we generated the nerve
 +
via our particle model, and Expert indicates nerves generated from the smoothed expert segmentation
 +
of real nerves. Right: Summary of the 12 segmentation results on real patient data. The red bars indicate the
 +
medians of distance between the automatic and the expert segmentation; the blue bars indicate the
 +
90th percentile, in voxels.]]
 +
 
 +
[[File:NerveSegResults3D.png|800px|thumb|center|
 +
Patient data results. Left: Rendered segmentation results a herniated disk (manually
 +
segmented, yellow) is impinging on the nerve tracts (segmented with our algorithm with minimal user
 +
input, green). Right: Slice that includes a section of a nerve and a ganglion. The yellow outline
 +
indicates the automatic delineation of the nerve. The algorithm tends to slightly under-segment the
 +
nerve, as we see the margins of the nerve underneath the segmentation.]]
 +
 
 +
We evaluate our algorithm on synthetic and patient data. We show that it can
 +
fully segment the nerves from the initial input point up to the end of the ganglia,
 +
and provide accurate estimates of the nerve thickness. To quantify the accuracy of our
 +
method, we measure the distances between the desired and automatic nerve surfaces.
 +
 
 +
First, we generate synthetic nerve tracks via our particle model. We vary the dynamics
 +
parameters, such as the radius and control points, beyond ranges that are observed in
 +
real images, to allow for tracks with more irregular behavior. The image intensity is
 +
then formed by adding white noise to the predicted nerve image. We superimpose the nerve
 +
tracks on backgrounds with no noise (blank), Gaussian noise, Perlin noise, and random sections of MRI volumes, which present varying
 +
degrees of segmentation dificulty.
 +
 
 +
We evaluate the algorithm on 20 images with each background. Typical diameters measure between
 +
four and ten voxels, yielding many partial volume voxels between nerve and background.
 +
We also construct and test a second synthetic dataset of the same size, where the
 +
initial nerve intensity is generated from smoothed binary map of expert nerve
 +
segmentation in patient data. The nerve and background noise models are applied in
 +
the same manner as in the first synthetic set. The nerves range from four to 25 voxels
 +
in diameter, and present with far more irregular shape.
 +
The median distance between the automatically extracted and true nerve surfaces for each nerve generated from our particle model is
 +
0.8±0.4 voxels, with a 90th percentile of 2.1±0.7 voxels. Most of the algorithm errors are in the edge, partial-volume voxels. For nerves generated from expert segmentations,
 +
the shape irregularity results in an increased median distance between the automatically extracted and true nerve surfaces of 1.3 ± 0.3 voxels, with the 90th percentile of
 +
4.1 ± 0.4 voxels.
 +
 
 +
 
 +
 
 +
We further demonstrate our method on MRI spine scans of 12 nerve bundle segmen-
 +
tations from six subjects. The scans were acquired 3D Wide-band Steady State Free
 +
Precession sequence. These
 +
include four nerves in two pathologies where the nerves have been displaced by disk
 +
herniations. We obtain both expert and automatic segmentations of nerve bundles inside
 +
the spine and ganglia that were deemed traceable, and evaluate tracing accuracy
 +
for all bundles. Figure 3 illustrates an example automatic segmentation.
 +
Quantitative results are summarized in Figure 2. The nerve diameters range from
 +
three to six voxels inside the spine and up to about 25 voxels in the ganglia. We find
 +
that the median distance between the automatically extracted and the expert surface
 +
is 1.0 voxel for most nerves, and the 90th percentile is 2.9 ± 0.6 voxels.
  
 
= Conclusion =
 
= Conclusion =
Line 29: Line 82:
 
As shown in the results, the proposed segmentation may slightly over-segment (usually by at most two voxels) in thin areas and under-segment in thick areas, but will give a very good estimation of the nerve core and location.
 
As shown in the results, the proposed segmentation may slightly over-segment (usually by at most two voxels) in thin areas and under-segment in thick areas, but will give a very good estimation of the nerve core and location.
  
 +
=Key Investigators=
 +
* MIT: Adrian Dalca, Polina Golland
 +
* BWH: Giovanna Danagoulian, Ehud Schmidt, Ron Kikinis
  
==Key Investigators==
+
= Publications =
* MIT: Adrian Dalca, Polina Golland
+
*[http://www.na-mic.org/publications/pages/display?search=Projects%3ANerveSegmentation&submit=Search&words=all&title=checked&keywords=checked&authors=checked&abstract=checked&sponsors=checked&searchbytag=checked| NA-MIC Publications Database on nerve segmentation]
* BWH: Giovanna Danagoulian, Ehud Schmidt
 
* SPL: Ron Kikinis
 

Latest revision as of 19:44, 28 November 2012

Home < Projects:NerveSegmentation
Back to NA-MIC Collaborations, MIT Algorithms,

Nerve Segmentation

Automatic segmentation of neural tracts in the dural sac and outside of the spinal canal is important for diagnosis and surgical planning. The variability in intensity, contrast, shape and direction of nerves in high resolution MR images makes segmentation a challenging task.

In this project, we present an automatic tracking method for nerve segmentation based on particle filters. We develop a novel approach to particle representation and dynamics, based on Bezier splines. Moreover, we introduce a robust image likelihood model that enables delineation of nerve bundles and ganglia from the surrounding anatomical structures. In the results section below, we summarize accurate and fast nerve tracking when compared to expert manual segmentation.

Data

We show four slices of an example High-Resolution MR scan. The blue outline in the first figure shows a manual segmentation of a nerve for reference. All arrows point to the same nerve bundle. Blue arrows show examples of poor contrast between the nerve and the surrounding tissue; orange arrows indicate the thickening of the neural tract into a ganglion.

Description

We present a tracking approach based on particle filtering, also known as sequential Monte Carlo tracking. Tracking has also been used successfully for segmentation of tubular structures. Most vessel tracking methods model the state as a cross-sectional ellipse or as a cylindroid. In tracking nerve bundles, the regions of low contrast require the state to capture substantially longer segments of the track than what is represented by a cross-section. In addition, nerves tend to change direction, often sharply, which necessitates a use of more complex descriptors than cylinders. Several of the data features can be seen in the Data section above.

To address the challenges of nerve tracking, we define a flexible particle representation that captures the geometric behavior of the nerve bundles. We use a Bezier spline centerline with a linear radius function to characterize a nerve bundle. We devise a dynamics model for particle updates that encourages continuity and smoothness. Furthermore, we define an image likelihood model that compares gradient fields and intensities of predicted patches with image observations to evaluate a posterior distribution of the particles' importance. With the particle model, dynamics model and likelihood model, we can implement a full particle filter. Once tracking is completed, we remove spurious segmentations by measuring the quality of the entire tract.

To the best of our knowledge, this method is the first automatic segmentation of nerve bundles and ganglia.

Results

Left: Summary of synthetic results for the eight types of synthetic nerves and backgrounds. The red bars indicate the average of the median distance between the automatic and the manual outline. The blue bars indicate the 90th percentile. For each category, Model indicates that we generated the nerve via our particle model, and Expert indicates nerves generated from the smoothed expert segmentation of real nerves. Right: Summary of the 12 segmentation results on real patient data. The red bars indicate the medians of distance between the automatic and the expert segmentation; the blue bars indicate the 90th percentile, in voxels.
Patient data results. Left: Rendered segmentation results a herniated disk (manually segmented, yellow) is impinging on the nerve tracts (segmented with our algorithm with minimal user input, green). Right: Slice that includes a section of a nerve and a ganglion. The yellow outline indicates the automatic delineation of the nerve. The algorithm tends to slightly under-segment the nerve, as we see the margins of the nerve underneath the segmentation.

We evaluate our algorithm on synthetic and patient data. We show that it can fully segment the nerves from the initial input point up to the end of the ganglia, and provide accurate estimates of the nerve thickness. To quantify the accuracy of our method, we measure the distances between the desired and automatic nerve surfaces.

First, we generate synthetic nerve tracks via our particle model. We vary the dynamics parameters, such as the radius and control points, beyond ranges that are observed in real images, to allow for tracks with more irregular behavior. The image intensity is then formed by adding white noise to the predicted nerve image. We superimpose the nerve tracks on backgrounds with no noise (blank), Gaussian noise, Perlin noise, and random sections of MRI volumes, which present varying degrees of segmentation dificulty.

We evaluate the algorithm on 20 images with each background. Typical diameters measure between four and ten voxels, yielding many partial volume voxels between nerve and background. We also construct and test a second synthetic dataset of the same size, where the initial nerve intensity is generated from smoothed binary map of expert nerve segmentation in patient data. The nerve and background noise models are applied in the same manner as in the first synthetic set. The nerves range from four to 25 voxels in diameter, and present with far more irregular shape. The median distance between the automatically extracted and true nerve surfaces for each nerve generated from our particle model is 0.8±0.4 voxels, with a 90th percentile of 2.1±0.7 voxels. Most of the algorithm errors are in the edge, partial-volume voxels. For nerves generated from expert segmentations, the shape irregularity results in an increased median distance between the automatically extracted and true nerve surfaces of 1.3 ± 0.3 voxels, with the 90th percentile of 4.1 ± 0.4 voxels.


We further demonstrate our method on MRI spine scans of 12 nerve bundle segmen- tations from six subjects. The scans were acquired 3D Wide-band Steady State Free Precession sequence. These include four nerves in two pathologies where the nerves have been displaced by disk herniations. We obtain both expert and automatic segmentations of nerve bundles inside the spine and ganglia that were deemed traceable, and evaluate tracing accuracy for all bundles. Figure 3 illustrates an example automatic segmentation. Quantitative results are summarized in Figure 2. The nerve diameters range from three to six voxels inside the spine and up to about 25 voxels in the ganglia. We find that the median distance between the automatically extracted and the expert surface is 1.0 voxel for most nerves, and the 90th percentile is 2.9 ± 0.6 voxels.

Conclusion

As shown in the results, the proposed segmentation may slightly over-segment (usually by at most two voxels) in thin areas and under-segment in thick areas, but will give a very good estimation of the nerve core and location.

Key Investigators

  • MIT: Adrian Dalca, Polina Golland
  • BWH: Giovanna Danagoulian, Ehud Schmidt, Ron Kikinis

Publications