This poster is published under an
open license. Please read the
disclaimer for further details.
Keywords:
Breast, Mammography, Observer performance, Cancer
Authors:
N. Karssemeijer1, R. Hupse1, G. J. den Heeten1, M. Lobbes2, R. M. Mann1; 1Nijmegen/NL, 2Maastricht, LB/NL
DOI:
10.1594/ecr2014/C-0358
Methods and materials
Image Dataset
200 full-field digital mammography cases (all from Hologic Selenia Systems),
acquired between 2003 and 2008 from the Preventicon screening centre in Utrecht (the Netherlands),
were used in this retrospective reader study.
All images were anonymized and institutional review board approval was waived.
Only cases with bilateral craniocaudal (CC) and mediolateral oblique (MLO) views were selected.
The final dataset consisted of 63 randomly selected screen-detected cancers,
17 cancers that were retrospectively judged to be visible,
20 false-positive referrals,
and 100 non-cancer cases.
Interpretation of this dataset was more challenging than for a typical screening population,
as obvious lesions and obvious benign abnormalities were excluded,
as well as cases with only microcalcifications present.
Mammogram Interpretation
12 readers participated in the study; nine were certified breast screening radiologists with 1-24 years of experience and 3 were residents.
Readers were informed about the approximate proportion of abnormal cases and that no microcalcifications were included in the dataset,
and also had access to prior mammograms for comparison (digital or film).
After a short training session on the features of the display station (e.g.
zooming and contrast),
readers were asked to identify all potential abnormalities.
Readers marked the location of each suspicious finding,
assigned a suspiciousness score (0-100),
and noted if they would refer the case or not in normal screening practice.
Analysis of sensitivity by number of readers and breast density
Breast cancer localization performance was determined by correlating and averaging reader scores corresponding to the same findings.
Based on the suspiciousness scores given for each lesion,
“positive” thresholds were set,
such that 10-20% of non-cancer cases were referred as positive by the reader(s) (i.e.
fixed false positive rate of 0.1-0.2).
Cancer location was determined to be correct if the distance to the center of the mass was within 2 cm.
The mean True Positive Fraction (TPF) for single,
double and multiple readers,
was determined and stratified by volumetric breast density (VBD).
VBD was determined for each case using automated software (VolparaDensity™,
version 1.4,
Matakina Technology,
Wellington NZ),
with cases designated as dense (>10% VBD; 60 cases) or non-dense (< 10% VBD; 140 cases).
The 10% threshold is approximately a mid-BIRADS 3 breast density category.
It was chosen in order to obtain sufficient numbers of dense cases for the analysis.