Академический Документы
Профессиональный Документы
Культура Документы
I. INTRODUCTION
Manuscript received December 11, 2012; revised January 18, 2013; accepted
January 20, 2013. Date of publication January 25, 2013; date of current version
April 27, 2013. This work was supported within the NeoMark project (FP7ICT-2007-2-224483) by the European Commission. Asterisk indicates corresponding author.
*S. Steger is with Fraunhofer IGD, 64283 Darmstadt, Germany (e-mail:
sebastian.steger@igd.fraunhofer.de).
N. Bozoglu, A. Kuijper, and S. Wesarg are with Fraunhofer IGD, 64283
Darmstadt, Germany.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TMI.2013.2242901
Fig. 1. Radial ray based segmentation methods find a closed surface around a
seed point by assigning a radius to each radial ray.
STEGER et al.: APPLICATION OF RADIAL RAY BASED SEGMENTATION TO CERVICAL LYMPH NODES IN CT IMAGES
889
Fig. 2. Examples of LNs in CT images illustrating the challenges in LN segmentation. The gold standard segmentation is shown in green. (a) Typical healthy
bean shaped LN embedded in fat. (b) Enlarged LN with central necrotic tissue. (c) LN close to a bone in noisy CT image. (d) LN with surrounding tissue of similar
density.
going. A small part of the existing methods target MRI [5], [6]
and Ultrasound [7] images, but the majority targets 2D CT slices
or CT volumes.
In 1996, Rogowska et al. compared in [8] state-of-the-art
methods including automatic thresholding, snakes and the watershed transform for the segmentation of LNs in 2D-CT slices
concluding that advanced image segmentation techniques are
required.
One year later, Honea et al. [9] presented a snake based
method tailored to LN segmentation in CT slices. The authors
concluded that a 3D active contour may overcome the issues
resulting from the slice by slice approach and, later on, presented such a method [10]. An initially small spheric model
is manually placed inside the LN. Simulated annealing with a
bias towards the LN exterior is used to accept/reject modified
contour points based on the combination of shape and appearance energy. The latter is based on the image gradient strength
weighted by the gradient direction. Unfortunately, the method
was only evaluated on artificial data.
Yan et al. [11] use level sets for the slice by slice segmentation of LNs in CT images. The external energy depends on the
image gradient and boundary leaking is prevented by a manually placed sphere serving as hard stopping criteria.
Lu et al. presented in [12] an interactive slice by slice segmentation approach based on live-wires. Alternatively, points
on the contour guiding the live-wire can be detected automatically by casting radial rays from a seed point and selecting the
points with the largest image gradient magnitude for each ray.
The LN detection and segmentation method of Feulner et al.
[13] uses graph cut on a voxel grid for the segmentation part.
The graphs edge weights are computed from (joint) histograms
of the LN interior (boundary) extracted from manually annotated data. Further efforts are made to prefer spherical cuts.
Instead of local shape knowledge (e.g., smoothness), other
methods use an ellipse as global shape knowledge. Examples are
the joint segmentation and registration framework of Unal et al.
[5] for LNs in MRI slices and the method of Zhang et al. [7],
which uses an ellipse as a global shape prior in a graph-cut based
approach for the segmentation of LNs in ultrasound slices. In
both algorithms, a user defined rectangle serves as initialization.
Furthermore, Moltz et al. [14] use an ellipse extracted from a region growing result for setting markers of a watershed transform
in a multi-step 3D LN segmentation method working on CT
890
Fig. 3. Distribution of CT-intensities of LNs with and without necrosis including and excluding a 1-mm margin at the boundary.
III. MATERIALS
The data set used in this work was built from 25 clinically
acquired contrast enhanced head&neck CT-scans with a slice
thickness between 0.5 and 1.2 mm from 25 patients suffering
from oral cancer. A total of 100 lymph nodes was selected and
a gold standard segmentation was created for each of them. This
set is divided into two subsets (set 1 and set 2).
The 49 LNs in set 1 were selected from 13 of the 25 CT-scans
by an experienced radiologist. Five more radiologists independently delineated each LN in each CT slice. Based on that,
the STAPLE-algorithm [27] was used to create a consensus
segmentation which serves as the gold standard segmentation
for this set. In this set, the minor axis ranges between 5.2 and
32.5 mm with an average standard deviation of
whereas the major axis ranges between 9.2 and 47.6 mm with an
average standard deviation of
mm. Fig. 3 shows
CT-intensity histograms separately for LNs with and without
necrotic tissue. For non-necrotic LNs, 95% of all intensities in
the LN interior (1 mm minimum distance to the boundary) are
within an interval of 43 and 119 HU. The average is at 80 HU.
For set 2, a trained image analyst selected and manually delineated 51 LNs in the remaining 12 CT-scans. This manual segmentation serves as the gold standard for this set.
The algorithm design and most of the evaluation was carried
out solely on set 1 because 1) its gold standard is more reliable
and 2) during LN selection particular attention has been paid
to include cases reflecting the previously described challenges
(28.57% of the LNs contain central necrotic tissue, 10.2% show
very low contrast at the boundary and in another 10.2% of the
cases the CT images are very noisy). The purpose of set 2 is to
independently support the results achieved on set 1.
To compare two segmentations, we use a single quality index
[28], which is based on the voxel set based measures dice similarity coefficient (DSC) and volumetric overlap
error (VOE) as well as the surface distance based methods
average symmetric surface distance (ASD) and maximum symmetric surface distance (MSD). Whereas
indicates
perfect segmentation, a value of 75 corresponds to average
manual expert segmentation quality which has been computed
based on the five candidate segmentations for each LN in set 1
using leave-one-out cross-validation.
STEGER et al.: APPLICATION OF RADIAL RAY BASED SEGMENTATION TO CERVICAL LYMPH NODES IN CT IMAGES
(1)
891
(3)
With a specific appearance cost function, this approach has been
pursued by Barbu et al. in their LN segmentation framework
[22]. Since GD is a local optimization strategy, it is prone to
being trapped in a local minimum around the initialization.
2) Belief-Propagation (BP): Since the energy given by (1)
depends on a unary term and pairwise connections between direct neighbors only, the energy minimization problem is suited
to be minimized by the belief-propagation algorithm [29]. Messages are iteratively passed from nodes to neighboring nodes,
thereby globally inferring the most likely assignment. Although
optimal solutions are not guaranteed for loopy graphs, it has empirically shown good results for loopy graphs [30] and has ever
since been successfully applied to inference problems in many
different application scenarios. In our case, we use the algorithm
to jointly infer the most likely label for each ray . Messages
are passed between neighboring rays. Like in the GD based algorithm, shape knowledge is incorporated by demanding similar
labels for neighboring rays using (2).
3) Graph-Cut (GC): Another optimization scheme that is
suitable to find an optimal closed surface around a seed point
is graph-cut. The approach was described in [31] for general
surfaces and later applied in a radial ray based segmentation
framework for the segmentation of LNs [24] and gliomas [32].
The underlying vertex weighted directed graph consists of
one vertex
for each node on each ray . The weights
of the vertices are given by differences of the appearance costs
(4)
where
assigns costs to a specific point based on the appearance in the image ,
is the set of all rays adjacent to
ray ,
assigns costs to the labels of an adjacent pair of rays
based on local shape knowledge (e.g., smoothness) and balances the appearance and shape terms.
A. Optimization Techniques and Shape Knowledge
1) Gradient Descent (GD): Demanding that neighboring
rays
have as similar labels as possible is commonly used to
incorporate shape knowledge into a radial ray based segmentation framework. This can be expressed by
(2)
where the parameter controls the stiffness. This shape cost
prefers spherical structures.
Minimizing (1) with a gradient descent optimization scheme
is simple and fast. Starting from an initial configuration , the
optimizer iterates into the direction given by the energy gradient
until convergence is reached. Whereas the partial derivatives of the shape term are determined analytically, finite differences are used to approximate the partial derivatives of the ap-
892
otherwise
(6)
The interval
is set to
. It is centered
at the mean of typical LN intensities and generously includes
the entire LN intensity range. Now, LNs are the objects with the
highest image intensity and image gradients at their boundary
always point inwards [see Fig. 4(c)].
In the following, we will describe three image gradient and
three image intensity based cost functions suitable for the segmentation of LNs (Fig. 5).
1) Logarithmic Absolute Gradient Strength (LAG): The first
and simplest image gradient based cost function solely depends
on its magnitude
(7)
where
is the gradient operator, the convolution operator,
and
is a Gaussian kernel of the variance . The maximum
cost of
is chosen so that
is always
true. The logarithm ensures that strong gradients of surrounding
objects do not result in significant lower costs than weak gradients at the LN boundary.
2) Direction Weighted Gradient (DWG): Besides the magnitude, the direction of the image gradient has an impact on the
likelihood of a position belonging to the boundary. It always
points towards the interior of the LN. Gradients that are tangential to the LN boundary or even point outwards (e.g., the
transition from central necrotic tissue to LN tissue) do likely
not belong to the LN boundary and therefore will be rejected.
The cost function given by (7) is not able to consider this. Assuming somewhat spherical objects around the seed point, the
approximate inward direction is given by the radial ray and (7)
is extended
(8)
otherwise
is the image gradient and
is the direction of the radial ray. Since the direction of the surface normal is only approximately given by ,
the parameter is introduced to model this uncertainty as it has
already been done in [10].
3) Local Gradient Maxima (LGM): Despite the logarithmic
scaling, both image gradient based cost functions given by (7)
and (8) still assign lower costs to locations with a higher image
gradient magnitude although there is no justification for that. In
[24], an interesting solution to this is presented: Gradient costs
depend on the distance to the local image gradient maxima. At
those local maxima, the cost is independent from the gradient
magnitude. We adopt this idea and express the third image gradient based cost function by
where
(9)
contains all
where models a spatial uncertainty. The set
local maxima of the direction weighted image gradient along
the radial ray through
(10)
STEGER et al.: APPLICATION OF RADIAL RAY BASED SEGMENTATION TO CERVICAL LYMPH NODES IN CT IMAGES
893
Fig. 5. Different image gradient (the three on the left) and image intensity (the three on the right) based appearance costs for a specific lymph node. (a) Log. absolute gradient strength (LAG). (b) Direction weighted gradient (DWG). (c) Local gradient maxima (LGM). (d) Homogeneity (HOM). (e) Distance from estimated
boundary (DST). (f) Accumulated intensities (ACI).
where
is the radius of the th bin and
is the
weighted image gradient at the th point of the radial ray with
the direction
. The weights are set based on the direction
of the image gradient analogously to (8).
4) Homogeneity (HOM): All previously described cost
functions only depend on the discontinuity at the LN boundary.
Properties of the LN interior, however, are not considered.
One exploitable property is the gray value homogeneity in the
interior, as it has for example been done in [24]. We express the
homogeneity for the th point on a ray by
(14)
otherwise
(11)
where
is the average intensity and
is a Gaussian kernel suppressing negligible inhomogeneities
possibly resulting from noise. Low costs are assigned to points
that show a homogeneous gray value profile between them and
the seed point. The logarithm ensures high granularity for small
inhomogeneities while, at the same time, covering a large range
of homogeneities. This cost function does not explicitly consider central necrotic tissue.
5) Distance From Estimated Boundary (DST): Alternatively, gray value intensities in the LN interior can be considered
based on the known absolute intensities on the Hounsfield scale.
Along the radial ray, we first search for the closest bin to
the center that does not fulfill a certain image intensity based
criterion
(12)
where is the index of the first bin whose intensity
is
within the interval
. In case of the presence of
central necrotic tissue around the seed point, corresponds to
the first bin outside the central necrotic and inside the regular
LN tissue. This property enables the support of central necrotic
tissue. The cost at a point on that ray depends on its distance to
the estimated boundary given by
(13)
This cost function is almost identical to the one presented in
[23]. The main difference is the determination of the estimated
boundary , where only the gray value difference to the intensity
at the seed point is considered.
894
TABLE I
FOR ALL POSSIBLE
AVERAGE SEGMENTATION QUALITY
COMBINATION OF IMAGE GRADIENT AND IMAGE INTENSITY
BASED APPEARANCE COST FUNCTIONS
for all three methods (GD, BP, GC) reported in Section IV-A.
To reduce other influences, most parameters are kept constant
(
,
). The cost function combination is expressed
by
(16)
(or
for GC, which does not have the parameter
) is maximized using the downhill simplex optimizer for all
three training sets and for all nine possible combinations. Following the idea of cross-validation, the segmentation quality
has then been computed on the corresponding test sets. The resulting average segmentation quality is depicted in Table I. The
best result (
) is achieved when combining DWG with
ACI using GC, but also other combinations are suitable.
Overall Parameter Optimization: Based on the cost function selection of the previous step, we determine the optimal
parameters for the different segmentation techniques that have
previously been regarded as constants. This is again done by a
downhill simplex based optimization with
as the target
function for all three training sets and the three segmentation
methods. This time, the variable segmentation parameter vector
consists of the number of rays as well as the shape cost
parameters and for GD and BP or for GC, respectively.
A detailed analysis of the achieved segmentation quality can be
found in Section VI.
Since combining DWG with ACI using GC results in the
best segmentation accuracy, the whole parameter optimization
process was executed again for this configuration, this time on
the entire set 1, resulting in the final parameters recommended
in a clinical setting (see Table II).
V. EXTENSIONS
The presented method has the following undesired properties.
The segmentation result depends on the seed point and thus
on the operator.
The shape cost expressed by (2) is minimal for spheres.
LNs, however, are often not-that-spherical. A more suitable shape cost may increase segmentation accuracy.
Low image contrast, noise, and other effects may lead to a
segmentation not expected by the user. A convenient way
to interactively manipulate the result is desirable.
We now present three extensions to radial ray based segmentation to address these issues.
STEGER et al.: APPLICATION OF RADIAL RAY BASED SEGMENTATION TO CERVICAL LYMPH NODES IN CT IMAGES
895
TABLE II
OVERVIEW OF THE PARAMETERS OF THE RADIAL RAY BASED SEGMENTATION
FRAMEWORK RECOMMENDED FOR THE USE IN A CLINICAL SETTING AND
THEIR WAY OF SELECTION
Fig. 6. For a given point, the expected radius of a neighboring ray is defined
by the intersection of that ray with a plane through the point orthogonal to the
image gradient direction.
otherwise
(17)
Fig. 7. The interactive extension allows the user to correct the boundary.
(19)
896
an independent verification and for the comparison to other approaches. The centroid of the gold standard segmentation was
used as the seed point if not stated otherwise.
A. Segmentation Accuracy
1) Theoretical Limit: The supported set of shapes of any radial ray based segmentation technique is theoretically limited to
all objects for which each point within the object can be directly
connected to the seed point without intersecting the object surface. This includes all convex shapes, but also all star convex
shapes if the seed point is within the set of points with direct
connection to the entire object surface.
The applicability of radial ray based segmentation to LNs
is assessed in our first experiment. The best possible segmentation is created by removing all voxels from the gold standard segmentation that cannot be connected to the seed point
without intersecting the LN surface. This segmentation is then
compared to the original gold standard segmentation. Typically,
only very few voxels close to the LN boundary do not satisfy
the required condition and the achieved average segmentation
quality is
. Further statistics are depicted in
Fig. 8. Only for one single outlier, a significant part of the object cannot be represented by any radial ray based method with
the seed point in the centroid (
). Fortunately, moving
the seed point is able to recover this bad result and the resulting
segmentation quality for this LN becomes
.
To conclude, the LN shape is very unlikely to be the bottleneck of a radial ray based segmentation technique.
2) Optimization Approaches: The radial ray based segmentation methods with the different optimization techniques, along
with the corresponding kind of shape knowledge and the most
suitable cost function combination with the final parameters
depicted in Table II have been applied to the test sets. Then,
the similarity/error measures described in Section III have been
used to compare the segmentation results to the gold standard.
Results are summarized in Table III. Statistics about the segmentation quality for different types of LNs are shown in Fig. 8.
The examples for different types of LNs shown in Fig. 9 have
been selected based on the achieved segmentation quality of the
GC based approach (minimum, 0.25-quant, median, 0.75-quant,
maximum).
STEGER et al.: APPLICATION OF RADIAL RAY BASED SEGMENTATION TO CERVICAL LYMPH NODES IN CT IMAGES
AVERAGE
897
TABLE III
STANDARD DEVIATION OF THE SEGMENTATION QUALITY FOR DIFFERENT SEGMENTATION TECHNIQUES MEASURED ON 49 LNS
Fig. 9. Example results of the GD (yellow), BP (blue), and GC (red) based segmentation method, respectively, compared to gold standard (green) for different
types of LNs. Selection on the examples is based on the segmentation qualities of the GC approach.
mass spring model based LN segmentation method of Dornheim et al. [17], [20] to the sets to compare the method to a
different segmentation algorithm class. See Section II for a brief
description of those algorithms. The parameters given in the respective papers were used. The results depicted in Table III and
Fig. 11 show that the proposed methods outperform the state of
the art. While the approach of Wang et al. achieves good results
in most cases and especially for small rather spherical LNs, it is
not able to cope well with low contrast images and the enlarged
necrotic cases. The approach of Barbu et al. uses a very simple
cost function and has a very low complexity. The segmentation accuracy remains moderate. Our results suggest that radial
ray based segmentation using a global optimization technique
such as GC is more suitable than model based segmentation for
LNs. Although efforts have been made in [18] to reduce its size
898
TABLE IV
COMPARISON OF THE PROPOSED GC BASED METHOD TO THE
STATE-OF-THE-ART ON AN INDEPENDENT SET OF 51 LNS
Fig. 11. Comparison of the proposed GC based method to the state of the art
for different types of LNs.
Fig. 12. Average and standard deviation of the mean segmenation quality and
relative volume deviations for different relative distances of the seed point to
the LN centroid using a logarithmic scale. Up to a distance of 25% the proposed
method (GC-SPI) is more robust than manual segmentation.
Besides a reduction in image evaluation time by a radiologist, another advantage of automatic segmentation approaches
is its high robustness, i.e., the same input data always results
in the same segmentation result. This enables tracking object
properties (e.g., the volume) over time with high validity. The
methods presented in this work, however, are semi-automatic
because the seed point has to be set manually. For a fair comparison of all methods, all experiments have been conducted
with the seed point in the centroid of the gold standard segmentation up to now. In this section, we aim to assess the dependency of segmentation quality and robustness on the seed
point for the GC based method (baseline) and the seed point invariant extended GC based method (GC-SPI). To this end, we
used 66 equally distributed seed points with different relative
distances to the centroid for all 49 LNs. This relative distances
were 5%, 10%, 25%, 50%, 75%, and 100% of the LNs minor
axis. For a specific relative distance and LN, the segmentation
algorithm was applied for all 66 seed points and the segmentation quality was measured. The average of those values is an
indicator for the quality whereas the standard deviation is related to the robustness.
Up to a relative distance of 25% from the seed point to the
LN centroid, the segmentation quality is hardly affected for both
the GC (
) and the GC-SPI (
) approach. However, the
robustness decreases a lot faster for the former approach. For
GC-SPI, the median of the relative standard deviation of 0.56%,
0.55%, and 0.93% for 5%, 10%, and 25% distance from the centroid, respectively, stays below 1%. For the baseline approch
(GC), these values are 1.03%, 2.27%, and 5.54%. Larger relative distances result in bad results, both in terms of segmentation
quality and robustness. Nonetheless, GC-SPI performs significantly better (in the median case
,
, and
for
50%, 75%, and 100% distance from the centroid, respectively).
To measure the relevance for practical use, we asked a radiologist to click in the approximate center of each LN and then
measured the distance to the LN centroid. The obtained average
relative distance of 21.6% and a median of 15.7% stays well
below the soft 25% limit and indicates that the GC-SPI is ready
for clinical use.
A key application scenario is tracking of LN properties over
time. For this scenario with the volume as the LN property, we
now compare our semi-automatic approach with manual segmentation in terms of segmentation accuracy and robustness.
The relative standard deviation from all 66 segmentations, we
obtained for different relative distances for each LN, were computed. The same was done for the five manual segmentations
available for each LN. We set those values in relation with the
STEGER et al.: APPLICATION OF RADIAL RAY BASED SEGMENTATION TO CERVICAL LYMPH NODES IN CT IMAGES
TABLE V
RUNTIME PERFORMANCE OF THE DIFFERENT OPTIMIZATION STRATEGIES.
ALL DURATIONS ARE REPORTED IN SECONDS
segmentation accuracies reported earlier. Fig. 12 shows the average and the standard deviation of those values for the different
relative distances and for the manual segmentations.
Up to a relative distance of 25%, the segmentation quality of
both approaches is close to manual segmentation, but the robustness is significantly higher. Assuming that different observers
set the seed point within a 25% radius relative to the minor
axis, the GC-SPI achieves a volume inter-observer variability
of 5.9% on the average. For 10% relative distance, this value
decreases to 3.0%. Both values are considerably lower than the
15.9% achieved by the human experts.
C. Runtime Performance
To give a rough impression of the runtime, we measured the
duration of the segmentation process for our implementations
of the different optimization techniques. All experiments have
been conducted on an Intel Core i72600K processor with 3.40
GHz and eight cores. Belief-propagation was implemented optionally using multiple threads (BP-MT) and the implementation of Boykov et al. [37] of graph-cut was used. The results are
depicted in Table V.
As expected, the local GD based technique is a lot faster than
the global optimization techniques. Nonetheless, the global optimization techniques are able to terminate within very few seconds. Exploiting parallelism ensures that the BP based method
achieves a runtime similar to the GC based method.
VII. CONCLUSION
In this paper, we apply radial ray based segmentation to cervical LNs in CT images, thereby coping with high variance of
size, shape, and surrounding tissue, as well as low contrast on
the object boundary, heavy image noise, and central necrotic
tissue inside the LN. From a user-defined seed point, radial
rays are cast uniformly in all directions and an optimization
strategy is used to find the best radius for each ray, considering image appearance and local shape knowledge. A local
GD optimizer has been compared with the global BP and GC
based algorithms along with suitable shape knowledge. Different image gradient and image intensity based image appearance cost function have been motivated and described, their
optimal parameters have been determined by maximizing the
amount of gold standard surface being close to a cost functions
minimum, and finally their combinations have been compared.
Furthermore, we have presented extensions to increase the robustness, to better cope with not-that-spherical shapes, and to
enable interactive corrections.
Evaluation in terms of segmentation accuracy, seed point dependency, and runtime performance has been conducted on a
set of 49 LNs of different sizes and shapes, containing normal,
899
noisy, low contrast, and necrotic LNs. Three-fold cross-validation and evaluation on an additional set of another 51 LNs has
ensured the validity.
The main results of this work are as follows.
1) Radial ray based segmentation supports virtually all occurring LN shapes.
2) The image based cost function and its parameters have a
large impact on the segmentation quality.
3) GC achieves the best results working on a combination
of the direction weighted image gradient and accumulated
intensities outside a predefined interval as the appearance
cost function.
4) The average segmentation accuracy (dice coefficient: 0.82,
average symmetric surface distance: 0.41 mm) is close to
manual segmentation and considerably higher than state of
the art radial ray based and model based LN segmentation
techniques.
5) Noisy images and especially surrounding tissue of similar
density result in a segmentation accuracy penalty.
6) The interactive correction enables a user to modify inaccurate segmentations, resulting in a higher segmentation
accuracy than manual segmentation while only requiring
an average interaction time of 12.1 s.
7) The typical variation of manually selected seed points does
not affect the segmentation accuracy.
8) Automatically iteratively adapting the seed point increases
the robustness by a factor of about 5.
9) Assessing the volume of LNs using the proposed method
instead of manual segmentation decreases the average
inter-observer variability from 15.9% to less than 5.9%.
10) Our algorithm usually terminates within 5 s.
We argue that the presented radial ray based segmentation techniques are mature enough to be used in clinical practice.
Future work includes the improvement of the segmentation
quality for LNs surrounded by tissue with similar density. Combining the radial ray based technique with model based segmentation is a promising direction. Furthermore, it will be applied
to LNs in other areas of the body and other imaging modalities
like MRI, as well as to other somewhat spherical objects such
as different types of tumors.
ACKNOWLEDGMENT
The authors like to thank Dr. G. Chiari and his colleagues
from the Department of Clinical Sciences, Section of Radiology,
Parma University Hospital, Italy, as well as Dr. I. Rodriguez
Jimenez and her colleagues from the Radio-Diagnosis Department, Centro Oncologico MD Anderson International Spain for
providing the test images, selecting the lymph nodes and, most
importantly, creating the manual segmentations.
REFERENCES
[1] Oral Cancer Mestastasis, J. Myers, Ed.. New York: Springer, 2010.
[2] L. H. Schwartz, J. Bogaerts, R. Ford, L. Shankar, P. Therasse, S.
Gwyther, and E. A. Eisenhauer, Evaluation of lymph nodes with
RECIST 1.1, Eur. J. Cancer, vol. 45, no. 2, pp. 261267, 2009.
[3] S. Steger, F. Franco, N. Sverzellati, G. Chiari, and R. Colomer, 3d
assessment of lymph nodes vs. RECIST 1.1, Acad. Radiol., vol. 18,
no. 3, pp. 391394, 2011.
[4] Head and Neck Cancer Imaging, R. Hermans, Ed. New York:
Springer, 2006.
900