506 / 1082
S491
ESTRO 36 2017
_______________________________________________________________________________________________
available imaging software (MIM Vista, Cleveland OH). On
the test dataset the agreement between the manually
drawn gold standard contours and atlas-based auto-
segmented contours was measured with a Dice-
coefficient. To determine the impact of auto-segmented
contours on dosimetry calculations we determined the
mean radiation dose for the manual contours and the auto-
segmented contours for left-sided breast cancer patients.
Differences in dose between the two contours were
expressed with mean absolute errors.
Results
Within the test dataset the atlas-based auto-segmentation
approach accurately delineated the heart with a Dice-
coefficient of 0.87 ± 0.06 (mean ± standard deviation).
Auto-segmentation was much less accurate for the LAD
with a Dice-coefficient of 0.05 ± 0.06. Among left-sided
breast cancer patients the mean heart dose was 1.2 ± 0.9
Gy for the manually contoured heart, and 2.7 ± 0.9 Gy for
the manually contoured LAD. The auto-segmented mean
heart dose was similar to the manually contoured mean
heart dose, with a mean absolute error of 0.1 ± 0.2 Gy
(range 0.0 - 0.7 Gy). The auto-segmented mean LAD dose
differed moderately from the manual contoured mean LAD
dose, with a mean absolute error of 1.0 ± 1.2 Gy (range
0.0 – 1.7 Gy). There were no statistically significant
differences between the manual contours and the
automated-contours for either the whole heart (p=0.78 by
Wilcoxon-rank sum test), or the LAD (p=0.85).
Conclusion
This study demonstrates that atlas-based auto-
segmentation accurately delineates the whole heart,
though less accurately captures the LAD. The high
concordance in mean heart dose between the manual
contours and automated contours suggests that atlas-
based auto-segmented contours could play a role in
radiation treatment planning.
PO-0898 Automated segmentation for breast cancer
radiation therapy based on the ESTRO delineation
guideline.
A.R. Eldesoky
1,2
, E.S. Yates
3
, T.B. Nyeng
3
, M.S. Thomsen
3
,
H.M. Nielsen
1
, P. Poortmans
4
, C. Kirkove
5
, M. Krause
6,7
,
C. Kamby
8
, I. Mjaaland
9
, E.S. Blix
10,11
, I. Jensen
12
, M.
Berg
13
, E.L. Lorenzen
14,15
, Z. Taheri-Kadkhoda
16
, B.V.
Offersen
1
1
Aarhus University Hospital, oncology, Aarhus, Denmark
2
Mansoura University, Clinical Oncology and Nuclear
Medicine, Mansoura, Egypt
3
Aarhus University Hospital, Medical Physics, Aarhus,
Denmark
4
Radboud University Medical Center, Radiation Oncology,
Nijmegen, The Netherlands
5
Catholic University of Louvain, Radiation Oncology,
Louvain, Belgium
6
OncoRay- University Hospital Carl Gustav Carus-
Technische Universität Dresden- and Helmholtz-Zentrum
Dresden-Rossendorf, Radiation Oncology, Dresden,
Germany
7
German Cancer Consortium DKTK Dresden and German
Cancer Research Center DKFZ Heidelberg, Radiation
Oncology, Dresden, Germany
8
Rigshospitalet, Oncology, Copenhagen, Denmark
9
Stavanger University Hospital, Oncology, Stavanger,
Norway
10
University Hospital of North Norway, Oncology,
Tromsø, Norway
11
Institute of Medical Biology- UiT The Arctic University
of Norway, Immunology Research group, Tromsø, Norway
12
Aalborg University Hospital, Medical Physics, Aalborg,
Denmark
13
Hospital of Vejle, Medical Physics, Vejle, Denmark
14
University of Southern Denmark, Institute of Clinical
Research, Odense, Denmark
15
Odense University Hospital, Laboratory of Radiation
Physics, Odense, Denmark
16
Hospital of Næstved, Oncology, Næstved, Denmark
Purpose or Objective
To internally and externally validate an atlas based
automated segmentation (ABAS) tool for loco-regional
radiation therapy of early breast cancer based on the
ESTRO consensus guideline for target volume delineation.
Material and Methods
Structures of 60 patients manually delineated according to
the ESTRO consensus guideline were included in four
categorized multi-atlas libraries (based on laterality and
surgery) using MIM Maestro
™
software. These libraries were
used for automated segmentation of primary and nodal
target volumes and organs at risk. Internal Validation of
ABAS was done by comparing ABAS before and after
correction against a gold standard manual segmentation
(MS) in 50 patients from the local institution using Dice
Similarity Coefficient (DSC), Average Hausdorff Distance
(AHD), difference in volume (∆V) and time. External
validation was done by comparing ABAS without correction
against MS in 40 patients from other institutions using DSC
and AHD. In the internal validation phase MS and
correction of ABAS was performed by only one observer,
while in the external validation phase MS was performed
by multiple observers from 10 different institutions.
Results
ABAS reduced the time of MS before and after correction
by 93% and 32%, respectively.
In the internal validation
phase, ABAS showed high agreement with MS for lung,
heart, breast and humeral head, moderate agreement for
chest wall and axillary nodal levels and poor agreement
for inter-pectoral, internal mammary nodal regions and
left anterior descending coronary artery (Figure 1).
Correcting ABAS significantly improved all the parameters
defined as increased DSC and decreased AHD and ∆V.
Applying ABAS in an external group of patients with
different arm positions, immobilization techniques and
respiratory gating status showed comparable results
(Table 1).
Table 1.