2084A Support Vector Machine (SVM) Classifier Enables Prediction Of Optimal Set Up, Prone Versus Supine, In Left Breast Cancer Patients

zoomzurichAI and Robotics

Oct 16, 2013 (3 years and 11 months ago)

88 views

2084 A Support Vector Machine (SVM) Classifier Enables Prediction Of Optimal Set Up, Prone Versus Supine, In Left
Breast Cancer Patients
J. Chang
1
, X. Zhao
2
, E. Wong
2
, Y. Wang
2
, S. Lymberis
1
, S. Formenti
1

1
NYU Langone Medical Center, New York, NY,
2
Polytechnic Institute of NYU, New York, NY
Purpose/Objective(s): NYU 05-181 Protocol studied prone vs. supine treatment positions for breast radiotherapy in 400
(200 right and 200 left) breast cancer carriers. Each patient underwent computed tomography (CT) simulation and
planning in both positions and the trial showed that while most patients were best treated prone, in 15% of left breast
cancer patients supine set up better spared the heart. We studied whether a SVM classifier applied to prone CT images
could successfully predict which left-breast patient is preferentially treated supine.
Materials/Methods: In 78 consecutive left-breast patients planning prone CT set was de-identified and exported in
DICOM RT format. Contours of critical organs (heart, lung, chest wall) and planning target volumes (index breast and
tumor cavity) were examined. Three-dimensional shape features including volume, centroid, second-order moments,
elongation, orientation and shape of chest wall were analyzed, along with relative 3D distances between heart, breast,
lungs, and chest wall. Volumes of in field heart and lung were also obtained. A weighting of 1:10 (prone-treated: supine-
treated) was used when determining the soft margin hyper plane of the SVM classifier with a K-fold cross validation
procedure to test the performance of the SVM classifier as well as to find optimal parameters for the classifier.
Results: Among the 78 left-breast patients analyzed 62 were treated prone and 16 supine, based on a selection of the
position that would best spare the heart after planning both supine and prone. When the SVM classifier was tested
against only prone CTs, by using 10 geometric features it achieved a 69% sensitivity (correctly classified prone/treated
prone) and a 75% specificity (correctly classified supine/treated supine). Adding heart and lung in-field volumes increased
sensitivity and specificity to 74% and 94% respectively. These preliminary results suggested that a SVM applied to an
initial prone CT in left breast cancer patients would have selected ~25% of patients to receive a second unnecessary
supine CT scan (since supine CT planning would have revealed inferior) and would have missed ~1% who would have been
better treated supine.
Conclusions: Preliminary data support the refinement of a feature-based classification to predict the optimal treatment
position from prone CT scans. Weighted SVM can achieve a high specificity at the expense of sensitivity. A relatively low
(~70%) sensitivity is acceptable since the patients classified as supine will be confirmed by receiving a second supine CT.
Since in-field heart and lung are significant predicting factors, knowledge of beam placement in addition to geometric
features is essential to successfully predict the optimal treatment position.
Author Disclosure:
J. Chang, None; X. Zhao, None; E. Wong, None; Y. Wang, None; S. Lymberis, None; S. Formenti, None.