This poster is published under an
open license. Please read the
disclaimer for further details.
Keywords:
Breast, Ultrasound, Biopsy, Cancer
Authors:
A. T. Stavros1, P. Lavin1, D. Schoenfeld2, M. J. Ulissey1; 1San Antonio, TX/US, 2Boston, MA/US
DOI:
10.1594/ecr2014/C-0926
Methods and materials
A prospective series of 66 cases included 37 histologically proven cancers and 29 histologically proven benign masses.
Each patient had five specific features assessed by Imagio using a 0-5 ordinal scale.
The internal tumor features included density of vascularity (DV),
blood oxygen saturation (BO),
and the total blood accumulation (BA) while external tumor features included total blood (TB) and peri-tumoral radiating vessels (RV); each were scored on a 0-5 ordinal scale and were summed to get a total internal score,
total external score,
and a total score.
Five classification methods were used to classify between benign and malignant outcomes:
- Logistic Regression (LR)
- Support Vector Machines (SVM)
- Classification Trees (CT)
- Random Forests (RF),
and
- K-Nearest Neighbors (KNN).
Ten-fold cross validation was used where the 66 cases were randomly divided into 10 groups.
Each of the ten possible groups was removed in turn from the 66 observations and the classifier was trained on the remaining groups to develop a classification rule.
This rule was then applied to the removed group.
This process was repeated ten times until every observation was assigned by a classifier that had not previously been developed using that observation.
A two-sided exact Jonckheere-Terpstra test was used to test the relationship between increasing scores (internal,
external,
total) and higher cancer grade.