IEEE Consumer Electronics Magazine - January/February 2020 - 64

Smart Home Healthcare

Figure 3. Extraction result by 2D display. (a) Original image. (b) Region extraction result by graph cuts. (c) Region
extraction result by supervoxel and graph cuts. (d) Proposed method. The green region is the correct region, the red region
is the extracted region, and the yellow region is the correct region extraction region.

The process consists of the initial TS image
generation, GGOs region of interest (ROI) extraction, graph cuts-based GGOs 3D ROI detection,
support vector machine (SVM)-based shadow
removal,15 and then use of the coordinates of the
TS images to perform voxel matching to the previous image. The current image and the registered
image are subtracted to obtain the final result.

EXPERIMENTAL RESULTS AND
DISCUSSIONS
An image registration experiment is conducted on 20 cases. The experimental data were
the CT images taken with a Toshiba Aquilion that
contained one or more GGO shadows per case.
The final time-sequential difference images are
used to evaluate the accuracy of region extraction, discrimination, and the final difference
image over time. The processing time for each
case is about 30 min.
Three types of region extraction methods
were used: 1) a region extraction method with
only graph cuts; 2) a region extraction method
using supervoxel and graph cuts; and 3) the proposed supervoxel and graph cuts method using

the gradient vector concentration degree. The
comparison of accuracy by the region extraction
method was tested. The results were evaluated
by the true positive rate (TP) and the false positive rate (FP). In the case of only graph cuts, the
average TP ¼ 49.46% and FP ¼ 26.89%, the average TP ¼ 63.911% and FP ¼ 28.002% in the graph
cutsþsupervoxel method, and the average TP ¼
71.835% and FP ¼ 9.539% were obtained from the
proposed method. The correctly and manually
extracted data are used as the ground truth to
the calculated TP and FP as TP ¼ n(B)/n(A), and
FP ¼ n(D)/n(C), respectively. n(A) is the number
of correct answers of GGO, n(B) is the number
correctly identifying GGO, n(C) is a number of
regions other than GGO, and n(D) is a number
incorrectly identified as GGO. Figures 3 and 4
show some of the results of area extraction by
the 2D and 3D displays.
Figure 5 shows an example of a blood vessel
extraction result. We further used SVM to identify the GGOs using the above-processed results.
The total number of ROIs in all 20 cases is 482, of
which 16 are excluded from the GGO regions,
excluding 4 cases which could not be judged as
GGO from the initial candidate regions. At this

Figure 4. Extraction result by 3D display. (a) Original image. (b) Region extraction result by graph cuts. (c) Region
extraction result by supervoxel and graph cuts. (d) Proposed method.

64

IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - January/February 2020

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2020
