200,000 data points) was classified in 0.366 seconds
by our implementation. This facilitates future real-time usage.
Finally, we consider whether an automatic classification procedure such as this can be useful in reducing the amount of time taken to process images in a
clinical setting. In an initial experiment, we found
that cardiologists would spend approximately five
hours analyzing a section of a blood vessel. We then
created a tool (figure 8) with our classifier built in.
The screen of this graphic user interface (GUI) is
divided into two main regions. The leftmost region
contains the tools provided to the user. There, the
user can select which view is most informative, adjust
image contrast and/or window level, and so on. The
right region is the work area where the user can interact with any of the views, slide along the pullback to
focus on the cross section of interest, make measurements, create annotations, and more. The cardiologist would run the classifier for a new image and
then, using the interactive tools, analyze the results
and correct some of the errors in the predictions.
The process, which the cardiologist follows, can be
described by following the process used in order to
annotate, classify, validate, and clean classification
results as shown in figure 7. In this figure, the leftmost column shows cryo-images (Roy et al. 2009)
while the second column from the left shows the
IVOCT images. Using the annotation function of the
plaque analysis tool, the expert would annotate the
image pixels as belonging to either calcium, lipid,
fibrous, or something else (used during training). The
third column from the left shows the result of this
annotation. It shows a mask, the same size as the
image itself, that indicates the location of each
plaque using colors. The next step includes running
the classifier, the results of which are shown in the
fourth column from the left. These results after preprocessing to remove isolated artifact predictions are
presented to the cardiologist (rightmost column).
We found that this process took at most an hour, a
reduction of 80 percent. This effort reduction indicates that improving the tool (figure 8) will make it
deployable in the near future.
Conclusion and Future Work
In this article, we have discussed an important
emerging application: an automated approach for
early plaque detection in blood vessels. Our approach
analyzes IVOCT images to solve this task. Using a
carefully designed feature set, we show that an SVM
with an RBF kernel is a high-accuracy classifier for
this task. Our results are of significant impact on this
important problem (Wagstaff 2012) with implica-
tions for early diagnosis of cardiovascular disease.
Now, for the first time, to our knowledge, it is possi-
ble to perform complete plaque analysis automati-
cally, enabling not only treatment planning for
plaque modification in real time but also to provide
enough information to perform studies on the effects
of various treatments of vulnerable plaque as well as
offline assessment of drug and biologic therapeutics.
In future work we will develop a complete software
suite for automated plaque characterization, creating
Figure 6. ROC Curve for All Three Plaque Types.
Area under the curve (AUC) values are 0.9837, 0.9947, and 0.9959 for calcium, lipid, and fibrous, respectively.
0 0.2 0.4 0.6 0.8 1
False Positive Rate
Table 2. Accuracy Results for
Accuracy Median Acc.
Overall 90. 70 ± 8.28%
Calcium 92. 14 ± 10.74% 100%
Lipid 96. 40 ± 8.87% 100%
Fibrous 100% ± 0.0% 100%
Table 3. Accuracy Results for Cryo-images.
Our Approach Baseline
Overall 81.15% 69.4%
Calcium 97.62% 66.88%
Lipid 87.65% 67.07%
Fibrous 97.39% 77.95%
Other 77.96% 30.46%