Dexterra PPT template 061103

1

• Classification of cardiac related artifacts in
  dynamic contrast breast MRI

Keith Stegbauera, Justin Smithab, Tanya
Niemeyera, Chris Wooda aConfirma, Inc. 821
Kirkland Ave, Kirkland WA 98033 bFirst Hill
Diagnostic Imaging, 1001 Boylston Ave, Seattle WA
98104
INTRODUCTION Breast MRI has been shown in a vast
number of investigations to have a high
sensitivity and specificity in the detection of
breast cancer. A breast MR study consists of a
pre contrast baseline scan followed by the
injection of gadolinium and the acquisition of
several post contrast series. Areas of rapid
contrast enhancement are considered suspicious
for malignancy. However, some normal anatomy
will exhibit this, particularly the veins,
arteries and the heart. The removal of
extraneous enhancing structures not within the
breast will improve clarity of breast MR
images. This study was designed to build and
evaluate the performance of a novel spatial
classifier for detecting enhancing clusters
within the chest wall. The used of such a
classifier improves the appearance and utility of
automatically generated MIP images in dynamic
contrast breast MRI. MATERIALS AND METHODS Data
Collection Data was collected from 21 different
institutions. Each institution acquired several
dynamic contrast breast MR studies, and one study
was selected at random from each institution.
The resulting 21 studies were used as a training
set for a classifier. Each study consisted of,
minimally, a pre-contrast series, a series
acquired approximately 1.5 minutes after the
administration of contrast, and a series acquired
4 minutes after the administration of contrast.
Any combination of pulse sequence type TR, TE,
resolution, number of slices, acquisition
orientation and resolution deemed acceptable by
the referring site were included in the study.
Cluster Segmentation All series were registered
in 2 dimensions and had an Angiogenesis map
created using CADstream 3.0 software (Confirma,
Inc. Kirkland WA). Within the Angiogenesis
map, all connected clusters of pixels were
evaluated by a technician trained to determine if
the cluster was inside or outside the chest wall.
Clusters inside the chest wall were defined as
those that were medial and posterior with respect
to the innermost intercostals muscle, not
including the internal thoracic vein and artery
when visible as an enhancing cluster. Using
the information in the DICOM header, the location
of the centroid (with respect to isocenter in the
imager) of each cluster was computed. A list of
all clusters within the study was generated, and
each cluster was assigned a value of true
(indicating that the cluster was within the chest
wall) or false (indicating that the cluster was
outside the chest wall). This list of locations
was then used as the training set for the spatial
classifier. Classifier Construction The
classifier was constructed in Matlab 6.5 R13
(Mathworks, Natik MA.) A cost function was used
as the input to an minimizing search algorithm.
The goal of the cost function was to define the
coefficients of a parabola opening in the
posterior direction of the patient that
encompassed as many artifactual centroids as
possible. Its representation was the
3-dimensional parabola with coefficients
Where CN is the number of Correct Negative
clusters, TP is the number of True Positive
clusters, FN is the number of False Negative
clusters and FP is the number of False Positive
clusters. b) Scoring by cluster location
with respect to P(X,Z) and volume
(3)
Where ?V(XXi) represents the sum of the volume of
each classification type, computed by labeling
the centroid of a cluster as anterior or
posterior to the classification line, then
counting the number of pixels labeled by
CADStream and multiplying by the voxel size. c)
Scoring by the anterior/posterior distance of
each centroid from P(X,Z)
Figure 5 The resulting classification line
plotted on an image from a breast MR study.
(4)
Where ?S(XXi) represents the sum of the distance
of each classification type, measured as the
anterior/posterior distance from the
classification line to the centroid of the
cluster. The values of ?, ?, ?, and ? were
constant throughout the study and across
classification types. Experimentally, good
values for these were found to be

• Table 2

The coefficients aj of equation (1) that were
converged on by minimization of equations
(2),(3), and (4) were considered to by the
coefficients of the optimum spatial classifier.
RESULTS After obtaining the optimized
coefficients for equation (1) for each of the
three cost functions, the resulting classifier
was applied to the testing set. This set was
composed of 7 studies randomly drawn from the
same pool as the 21 studies in the training set.
This set consisted only of studies that were
not in the training set, and contained a total of
231 clusters of pixels. The performance of the
classifier when rated by volume classified
(Figure 3) was consistently better than the
performance when rated by the number of clusters
(Figure 4) in each measure of performance, with
the exception of t he volume weighted negative
predictive value (NPV) and the distance weighted
specificity.
Figure 6. An example of the large, connected mass
representing the enhancing heart on a dynamic
contrast breast MRI study.
CONCLUSIONS This fully automated classifier can
be used to automatically eliminate distracting
regions of high uptake when evaluating maximum
intensity projections of subtraction images
(Figure 7). Additionally, the classifier can
assist in removing irrelevant of artifactual
enhancing regions when further evaluating pixels
in automated CAD for dynamic contrast breast MRI
studies.
Figure 7. Example of original MIP images made
from subtractions of dynamic contrast breast MR,
and MIP images constructed with the cardiac
enhancement removed by this classifier. Note the
enhancing, irregular region that becomes apparent
in this orientation when the cardiac enhancement
is removed.
(1)
Where X represents the distance from isocenter
along the patients right/left axis, Z represents
the distance from isocenter along the patients
superior/inferior axis. P(X,Z) represents a
location along the patients anterior/posterior
axis. The classifier performs a 9-dimensional
optimization of a cost function to find the best
solution for the coefficients a through j. Cost
Function Design Cost functions were designed to
maximize the sensitivity and specificity of the
classifier to both volume and number of clusters.
The utility of maximizing the performance of the
classifier to the number of clusters is to
eliminate the largest number of enhancing
clusters within the chest cavity. For scoring,
each cluster was assigned one of the following
values in the cost function.

• The important feature of this classifier is its
  robustness across acquisition techniques. The
  algorithm is also strong with respect to multiple
  data acquisition techniques. The effect of more
  or less sensitive acquisition protocols on the
  cluster is to increase or decrease the volume of
  the cluster, we have observed little effect on
  the location of the centroids in most clusters.
• With respect to the misclassifications by the
  classifier, in the non-cardiac volumes classified
  as cardiac volumes occurred in only four clusters
  for the volume weighted classifier and in six
  clusters for the distance weighted classifier,
  two of which were the same misclassifications as
  mentioned previously. The false positive
  misclassifications were due to one of the
  following
• a) Motion outside the chest wall
• b) Small vessels in the ribs, or sternal
  arteries.
• c) The region was in an area of poor SNR for
  the coil, leading the technician to misclassify
  it as inside the chest wall.
• Also of interest were clusters within the chest
  wall as outside the chest wall was more common.
  These generally represented pulmonary arteries,
  or small vessels. While it is desirable to
  classify all of these as being within the chest
  wall, a clinician can easily identify these as
  being of little interest in clinical practice.
  Also, these small, dim clusters demonstrate
  little absolute enhancement, limiting their
  impact on the quality of the MIP images.

Figure 3 The performance of each cost function
within the classifier with respect to the total
volume of the clusters detected by the parabolic
classifier.
Figure 4 The performance of each cost function
within the classifier with respect to the number
of clusters detected on each side of the parabola.

• Table 1

Three types of cost functions were
implemented a) Scoring by cluster location
It is also noteworthy that the unweighted
classifier performed without any clusters being
falsely classified as cardiac when they were not
inside the chest wall. The classifier results
in a parabola that is associated in size and
shape with the chest wall, without mirroring it
exactly (Figure 5). The artifacts of interest,
however, are for the most part a large connected
mass representing the heart (Figure 6).

• 1) L. Nunes, Breast Imaging Interpretation
  Model, Radiology, Vol 203 No3, 833841 1997.
• 2) C. Kuhl, Dynamic Breast MR Imaging Are
  Signal Intensity Time Course Data Useful for
  Differential Diagnosis of Enhancing Lesions?,
  Radiology, Vol 211 No 1101-110 1999.
• L. Nunes, Optimal Post-Contrast Timing of Breast
  MR Image Acquisition for Architectural Feature
  Analysis, JMRI,
• 1642-50 2002.
• 4) B. Tomanek, Probe With Chest Shielding for
  Improved Breast MRI, Mag. Res. In Med.
  43917-920, 2000.
• 5) D. Wald, MR-guided fine needle aspiration of
  breast lesions initial experience., J. Comput
  Assist Tomogr. Jan-Feb 20(1)1-8, 1996

(2)