ESTRO 36 Abstract Book

S549

ESTRO 36 2017

_______________________________________________________________________________________________

was observed in area of Hippocampus (less than 5 mm).

Conclusion

: With complete radiological diagnosis, HA-

WBRT can be delivered with oncological safety. Proper

margin definition of HA in delineation is to be confirmed

with individual technique.

PO-1004 Machine learning methods for automated

OAR segmentation

P. Tegzes

, A. Rádics

, E. Csernai

, L. Ruskó

General Electric, Healthcare, Budapest, Hungary

General Electric, Healthcare, Szeged, Hungary

Purpose or Objective

Manual contouring of organs at risk can take significant

time. The aim of this project is to use machine learning to

develop fully automated algorithms to delineate various

organs in the head and neck region on CT images.

Material and Methods

Machine learning models were built based on 48 CT

sequences of the head and neck region with 5 manually

contoured organs from the Public Domain Database for

Computational Anatomy. Data were randomly separated

to 32 train, 8 cross-validation and 8 test sequences. Three

different machine learning models were combined to

achieve automated segmentation. The first step uses a

support vector machine classifier to separate patient

anatomy from all other objects (a). The second step

applies slice-based deep learning classification to detect

the bounding box around the organ of interest (b). The

final step achieves voxel-level classification based on a

fully connected neural net on the voxel intensities of

suitably selected neighboring voxels (c). Very similar

model architectures were trained for all the different

organs.

Results

The body contour detection has been

previously trained

on another dataset containing full-body images and

achieved an average accuracy of 96.6%. The mean error of

the bounding box edges was 3mm, the corresponding dice

scores ranged from 72% to 94% depending on the organ of

interest. The first results of the voxel level segmentation

gave average dice values of 38% to 77% depending on the

investigated organ, and several opportunities for further

fine-tuning have been identified.

Conclusion

Machine learning methods can be competitive with

standard image processing algorithms in the field of organ

segmentation.

PO-1005 Automatic segmentation of cardiac sub-

structures in the treatment of HL

C. Fiandra

, M. Levis

, F. Cadoni

, V. De Luca

, F.

Procacci

, A. Cannizzaro

, R. Ragona

, U. Ricardi

University of Torino, Oncology, Torino, Italy

Purpose or Objective

to validate, in the context of treatment of Hodgkin

Lymphoma, three commercial software solutions for atlas-

based segmentation of cardiac sub-structures

Material and Methods

25 patients were selected and then divided into two

groups: 15 patients will make up the personalized atlas

and 10 patients on which will be applied atlases created

in order to assess its quality. For the selection of patients,

the following inclusion criteria were selected: patients

with HL presentation of a mediastinal mass at the onset of

the disease and the availability of CT imaging with

contrast. Two expert physicians have delineated on the

diagnostic CT with contrast the selected 15 patients

cardiac structures: the heart as a whole, the four

chambers of the heart, the coronary artery and valvular

structures; which will compose the atlas. We use three

commercial solutions (Velocity AI, MIM and RayStation) in

order to compare their results; the structures delineated

by doctors on the 5 control patients will be compared with

those automatically drawn by atlases, through the

conformality function (Dice Index (DI)). In addition, the

atlases underwent a clinical evaluation of the involved

physicians: in particular it was asked to a Radiation

Oncologist to analyze contours made by the three

software on reference patients to evaluate the goodness

of the warp made from atlases than those performed by

him. Clinical judgments were recorded on a scale of

numerical values: 1 = poor; 2 = medium; 3 = good.

Results

in terms of statistical analysis, the data obtained from the

values of Dice Index were compared structure by structure

between the three platforms. The Figure 1 shows only

structures with a Dice Index more than 0.5 (right atrium,

left atrium, the heart, the left side wall, interventricular

septum, aortic valve, left ventricle and right ventricle).

The differences between the 3 software were calculated

and the structures delineated by MIM have more

frequently higher values of Dice Index, compared to those

of Velocity and RayStation, with respectively 0.03 to 0.01

p-value. Instead the difference between Velocity and

RayStation is not statistically significant (p-value = 0.8).

Regarding the evaluation of the Radiation Oncologist as

compared to DI, values show that RayStation is the

software that realizes contours more applicable in clinical

practice, with statistically significant differences from

Velocity and MIM, with p-value respectively of 0.038 and

0.046. While the difference between Velocity and MIM is

not statistically significant (p-value = 0.083).

Conclusion

In general we can say that the contours applied by atlases

are valid, even if not yet optimal and they may represent

a starting point for the step of contouring, useful to speed

up this process; based on the values of Dice Index

collected in this study, MIM performs better while

RayStation appears the most powerful software from a

clinical point of view thus obtaining contours more