Mammogram image quality as a potential contributor to disparities in breastcancer stage at diagnosis: an observational study

Abstract

Background

In an ongoing study of racial/ethnic disparities in breast cancer stage atdiagnosis, we consented patients to allow us to review their mammogramimages, in order to examine the potential role of mammogram image quality onthis disparity.

Methods

In a population-based study of urban breast cancer patients, a single breastimaging specialist (EC) performed a blinded review of the index mammogramthat prompted diagnostic follow-up, as well as recent prior mammogramsperformed approximately one or two years prior to the index mammogram. Sevenindicators of image quality were assessed on a five-point Likert scale,where 4 and 5 represented good and excellent quality. These included 3technologist-associated image quality (TAIQ) indicators (positioning,compression, sharpness), and 4 machine associated image quality (MAIQ)indicators (contrast, exposure, noise and artifacts). Results are based on494 images examined for 268 patients, including 225 prior images.

Results

Whereas MAIQ was generally high, TAIQ was more variable. In multivariablemodels of sociodemographic predictors of TAIQ, less income was associatedwith lower TAIQ (p < 0.05). Among prior mammograms, lowerTAIQ was subsequently associated with later stage at diagnosis, even afteradjusting for multiple patient and practice factors (OR = 0.80,95% CI: 0.65, 0.99).

Conclusions

Considerable gains could be made in terms of increasing image quality throughbetter positioning, compression and sharpness, gains that could impactsubsequent stage at diagnosis.

Background

In the United States there is evidence that non-Hispanic (nH) Black women are morelikely to die from breast cancer compared to their nH White counterparts, despitehaving a lower incidence of the disease. This mortality disparity is especially highin Chicago, where most recent available data suggests that nH Black women die frombreast cancer at a two thirds higher rate than nH Whites [1]. Despite current controversies regarding the timing and frequency ofscreening with mammography [2–6], it is generally recognized as effective in reducing morbidity andmortality from breast cancer [7, 8]. Despite reporting similar mammography utilization [9], Black and Hispanic women continue to be diagnosed at a later stage ofbreast cancer compared to Whites [10] and this later stage is at least partly responsible for the greaterbreast cancer mortality experienced by Black women in the United States as comparedto Whites.

Prior data from Chicago suggest that nH Black and Hispanic women were less likelythan nH Whites to obtain screening mammography at facilities with characteristicssuggesting high quality screening, which include academic facilities, facilitiesthat relied on breast imaging specialists, and facilities that offered digitalmammography [11]. These apparent disparities in the distribution of mammography practicecharacteristics might result in a disparity in the quality of the process ofmammography screening and diagnostic follow-up. Racial/ethnic or socioeconomicdisparities in quality of mammogram images, radiologic interpretation of mammograms,or timeliness of diagnostic follow-up and resolution of an abnormal mammogram mightbe mediated by mammography practice characteristics. The goals of the presentanalysis were to (1) examine whether better image quality was associated withearlier breast cancer stage at diagnosis, and (2) examine whether there existeddisparities in image quality by race/ethnicity or socioeconomic status.

Our hypothesis was that racial and ethnic minorities and women of lower socioeconomicstatus (less education and income, and lacking private health insurance) would tendto be screened at lower resource facilities. These might include facilities notsituated in academic medical centers, that relied to a lesser extent on radiologistand technologist specialists, and that tended to use analog as opposed to digitalmammography. Digital mammograms may tend to be of higher quality than analog images [12].

This unequal distribution of mammography practice characteristics might translateinto lower image quality for these women. We conceptualized image quality as havingtwo components, one related to the skill of the technologist and one relatedprimarily to mammography machine calibration. Lower image quality has beenassociated with interval breast cancer, i.e., breast cancer presenting throughsymptoms despite a recent normal-appearing screening mammogram [13]. Symptomatic breast cancer is considerably more likely to be later stagethan screen-detected breast cancer [14]; thus, image quality might be associated with stage at diagnosis andmight help to explain disparities in stage at diagnosis.

Materials and methods

Sample and procedure

Patients for this study were recruited from the parent study, “BreastCancer Care in Chicago”, details of which have been previously published [15]. Briefly, female patients were eligible if they were diagnosedbetween March 1, 2005 and February 31, 2008, diagnosed between 30 and 79 yearsof age, resided in Chicago, had a first primary in situ or invasive breastcancer, and self-identified as either non-Hispanic White, non-Hispanic Black orHispanic. All diagnosing facilities in the greater Chicago area(N = 56) were visited monthly by certified tumor registrars employedby the Illinois State Cancer Registry (ISCR) and all eligible newly diagnosedcases were ascertained. Participants completed a 90-minute interview that wasadministered either in English or Spanish as appropriate using computer-assistedpersonal interview procedures. The final interview response rate was 56%representing 989 completed interviews among eligible patients (397 nH White, 411nH Black, 181 Hispanic, response rates 51%, 59% and 66%, respectively) [16]. Upon completion of the interview, patients were asked to provideconsent to allow abstraction of their medical records for information pertainingto their breast cancer diagnosis, and asked to allow the study to obtainoriginal breast screening and diagnostic images for the mammography reviewsubstudy. Both the main study and mammogram review substudy were reviewed andapproved by the University of Illinois at Chicago Office for the Protection ofResearch Subjects.

Mammogram review substudy

Patients reporting either initial awareness of their breast cancer throughscreening mammography or initial awareness through symptoms despite a priormammogram within 2 years of detection were eligible for this substudy(N = 597). Of these, 369 (62%) consented to a review of theirmammogram and other breast images involved in their screening and diagnosis.Original mammograms, diagnostic follow-up images and corresponding reports wererequested from screening and diagnostic facilities. Often, multiple facilitieswere involved for a single patient. In all, we received 494 mammograms performedon 268 patients. Approximately 90% of mammograms were bilateral, standard fourview mammograms, while the remainders were unilateral mammograms.

A single breast imaging specialist (EC) performed a blinded review of mammograms(blinded to the original interpretation and all other subsequent screening anddiagnostic mammograms and results). All reviews were blinded to patient age,race/ethnicity and other sociodemographic characteristics. Seven indicators ofimage quality were assessed: positioning, compression, sharpness, contrast,exposure, noise and artifacts [13, 17]. Each was scored on a five-point Likert scale, where 4 and 5represented good and excellent quality, respectively, while 1 represented poorquality. The 273 participants were less likely than eligible nonparticipants tobe minority (46% vs. 64%, p < 0.0005) and more likely to reportsymptomatic discovery (43% vs. 31%, p = 0.002), but were similar onother characteristics.

Analysis variables

Variables for image quality

We defined a continuous measure of image quality that was a simple sum of the7 indicators, with a theoretical range of 7 (lowest quality) to 35 (highestquality). Results of analyses using this variable were similar to resultsusing the binary version described next, and therefore these results are notpresented. We defined a separate binary variable to indicate higher imagequality as those images that received a score of at least 4 (very good) onall seven indicators. Positioning, compression, and sharpness are affectedby the patient-technologist interaction and the skill of the technologist,whereas contrast, noise, exposure, and artifacts are primarily a function ofmammography machine calibration. Therefore, we defined two additionalmeasures of image quality. Higher technologist-associated image quality(TAIQ) was defined as receiving a score of at least 4 (very good) onpositioning, compression, and sharpness, and a variable summing these threemeasures was defined. Higher machine-associated image quality (MAIQ) wasdefined as receiving a score of at least 4 (very good) on contrast, noise,exposure, and artifacts, and a variable summing these four measures wasdefined.

Measures of race/ethnicity and socioeconomic disadvantage

Race and ethnicity were self-reported at interview. Ethnicity was defined asHispanic if the patient self-identified as Hispanic, or reported a LatinAmerican country of origin for herself or for both of her biological parents.Race and ethnicity were used to categorize patients as non-Hispanic White,non-Hispanic Black or Hispanic. Socioeconomic disadvantage was defined fromself-reported annual household income, educational attainment, and healthinsurance status. Income was reported at interview in categories of less than$10,000, $20,000, $30,000, $40,000, $50,000, $75,000, $100,000, $150,000,$200,000, and greater than $200,000. Annual household income analyzed as anordinal variable and also categorized for some analyses as not exceeding versusexceeding $30,000. Formal education was reported in years, and was analyzed bothas an ordinal variable and categorized as not exceeding a high-school degreeversus having some post-secondary education. Health insurance status wascategorized as lacking private health insurance versus having any private healthinsurance. Patients with Medigap or similar supplementary private healthinsurance were defined as privately insured.

Mammography practice characteristics

Individual mammograms were defined with respect to image type as either digitalor analog (film screen). Although image type is an individual attribute of themammogram itself, it indicates something about the availability of digitalmammography and therefore we group it here with practice characteristics. Allmammography facilities included in these analyses were accredited by theMammography Quality Standards Act (MQSA). Mammography practices were grouped byfacility type into: (a) public, (b) private non-academic, (c) private with anacademic-affiliation, or (d) private university-based hospital or medicalcenter. For some analyses we dichotomized this variable to indicate whether afacility was located within an academic hospital or medical center. Facilitydata on the numbers and types of mammography technicians and radiologists wereavailable from a prior mammography facility survey of Chicago performed duringthe study period [11]. Facilities reported the number of general radiologists and breastimaging specialists interpreting mammographic studies. A breast imagingspecialist was defined as a radiologist who dedicated at least 75% of his/herworking time on breast imaging, regardless of fellowship training. We defined avariable describing each facility’s reliance on breast imaging specialistsas none, mixed, or sole reliance on specialists. We defined an analogousvariable to categorize facilities by the extent of their reliance on dedicatedmammography technologists (mixed vs. sole).

Clinical variables

Mode of breast cancer detection was defined as asymptomatic if the patientreported initial awareness of the breast cancer through mammography or otherbreast imaging in the absence of any symptoms, otherwise mode of detection wasdefined as symptomatic. Stage at diagnosis was categorized using the AJCCcategories of 0, 1, 2, 3, and 4 ( http://www.cancerstaging.org/).

Statistical analyses

Statistical analyses were conducted using SAS version 9 (SAS Institute, Cary, NC)and Stata version 11 (Statacorp, College Station, Texas). We tabulated theobserved distribution for each of the seven image quality indicators andestimated polychoric correlations [18]. The percentage of higher image quality mammograms was tabulatedagainst patient characteristics and mammography practice characteristics, andwas repeated for higher technologist-associated image quality and highermachine-associated image quality. Due to the greater variability of TA imagequality, only these associations are presented. P-values were estimated inunivariable logistic regressions using the Huber-White sandwich estimator toadjust standard errors for clustering of images within patients.

Multivariable logistic regression of higher image quality

Due to the greater variability of TA image quality, we focused on these imagequality indicators in the analyses that follow. We conducted multivariablelogistic regression models in order to estimate associations with higher TAimage quality while using the Huber-White sandwich estimator to adjust standarderrors for clustering of images within patients. The first model (baselinemodel) included terms for age and age squared. Next, we added variables forrace/ethnicity, income, education and private insurance status together. Throughbackwards elimination procedures we removed those variables with a p-value>0.10 via Wald tests.

Image quality indicators as predictors of stage at diagnosis

After excluding the index films and using data from only the prior films(performed prior to any symptoms or diagnosis), we conducted multivariableordinal logistic regression to estimate associations for higher TA image qualitywith breast cancer stage at diagnosis. Each image quality indicator was modeledseparately while adjusting for age, education, income, private health insurancestatus, academic vs. other facility type and image type (analog vs. digital). Weestimated odds ratios from ordinal logistic regression models with robuststandard errors.

Results

Image quality

Results are based on 494 images examined for 268 patients. Very good or excellentscores were less frequent for positioning, compression and sharpness (61, 66,and 68%) than for contrast, exposure, artifacts and noise (86, 89, 90 and 96%,respectively) (Table  1). As anticipated, polychoriccorrelations between the seven mammography quality indicators were higher withintechnologist-associated indicators (mean 0.86, range 0.79-0.94), and higherwithin machine-associated indicators (mean 0.84, range 0.77-0.93), than betweentechnologist-associated and machine-associated indicators (mean 0.63, range0.51-0.72) (Table  2).

Table 1 Distribution of image quality indicators (N = 494images)

Table 2 Polychoric correlations between the seven mammography qualityindicators

Patient and practice characteristics predict lower image quality

Racial/ethnic and socioeconomic disadvantage were associated with lower TA imagequality (Table  3). The percentage of films thatreceived a score of 4 or 5 on all 3 TA indicators was greater for nH White thanminority patients (57% vs. 49%, p = 0.13), greater for patientsreporting higher vs. lower income (58% vs. 45%, p = 0.03), andgreater for patients with more than a high-school education than for those withless education (58% vs. 46%, p = 0.04), but did not vary by privatehealth insurance status (Table  3). Better imagequality was considerably more likely for images performed at hospital-based,academic facilities than at other types of facilities. The extent to whichfacilities relied on full-time mammography technologists did not seem to beassociated with image quality. On the other hand, better image quality wasconsiderably more likely for images performed at facilities that relied solelyon breast imaging specialists than at facilities that did not (Table  3). Digital mammograms were considerably more likely to bescored as high quality than analog images. Results were generally similar whenwe examined all 7 indicators as a group. There was little variation inmachine-associated image quality indicators by racial/ethnicity or socioeconomicstatus (results not shown).

Table 3 Distribution of patient and practice characteristics with highertechnologist-associated image quality

Multivariable models of higher image quality

We conducted multivariable logistic regression in order to examine the extent towhich race/ethnicity and socioeconomic status were associated with imagequality. When racial/ethnic and socioeconomic variables were modeled together inlogistic regression models of higher image quality, only income was retained(p-value for income = 0.001) while race/ethnicity, health insurancestatus and education were not retained in the final model (Table  4). Results were very similar when modeling higher imagequality based solely on technologist-associated indicators (p-value forincome = 0.001), and again when modeling higher image quality basedsolely on machine-associated indicators (p-value for income = 0.03);in each instance, income and only income was retained (results not shown).

Table 4 Multivariable nested logistic regression models of higher imagequality

TA image quality predicts stage at diagnosis

In order to examine the extent to which variation in image quality was associatedwith breast cancer stage at diagnosis, we conducted multivariable ordinallogistic regression models. Higher image quality across all seven indictorscombined was inversely associated with breast cancer stage at diagnosis (OR =0.91, 95% CI: 0.80, 1.03) (Table  5). Higher imagequality for technologist-associated indicators was associated with earlier stageat diagnosis, whereas higher image quality for machine-associated indicators wasgenerally not associated with stage at diagnosis (Table  5).

Table 5 Higher quality mammography imaging and breast cancer stage atdiagnosis (N = 210 images prior to the index image with completedata on covariates)

Discussion

In our population-based sample of urban breast cancer patients, lowertechnologist-associated image quality was associated with later breast cancer stageat diagnosis, and patients with lower income were less likely to obtain high qualitymammography imaging. Our results suggest that differences in mammogram image qualitymay be contributing to socioeconomic disparities in stage at diagnosis.

Most of the variation in image quality was due to quality indicators that would tendto be associated with the skill of the technologist in working with the patient toensure proper positioning, adequate compression, minimizing blur (maximizingsharpness), and performing the mammogram again if the quality of the image wassuboptimal. Mammography technologists who spend most or all of their time performingmammograms may produce higher quality mammography images compared to technologistswhose duties are split, although there is little if any literature on this topic. Weanticipated that image quality would be better at facilities that relied solely ondedicated mammography technologists, but we did not find any evidence to supportthis expectation in the current study. We did find that mammogram images fromacademic institutions and from institutions relying on breast imaging specialiststended to be of higher quality. It may be that mammography technologists who workalongside breast imaging specialists have greater opportunities to expand theirknowledge and expertise, or these settings may tend to hire more skilledtechnologists to begin with. A specialized, high volume breast radiologist may bebetter able to find cancers in images that are of lower quality than generalradiologist or one that reads fewer mammograms. A tendency for lower quality imagingat non-academic facilities, in combination with less expertise available to readthose images, may result in less early detection and later stage at diagnosis.

Prior studies have found that digital mammogram images are more likely to be judgedas being higher quality compared to analog images [12]. In the present study, our expert consistently scored digital images asbeing of higher quality across all seven image quality indicators. Since we wereunable to blind our expert as to whether an image being reviewed was analog ordigital, it is conceivable that a bias or preference towards digital images may haveresulted in an artificially higher image quality score for digital images, but thisseems unlikely. Nonetheless, after controlling for type of mammogram in ouranalyses, lower income remained associated with lower quality imaging, and lowertechnologist-associated image quality was in turn associated with later stagediagnosis. Therefore, our results do not appear to be due to variation in the use ofdigital versus analog mammography across the study sample.

The Mammography Quality Standards Act (MQSA) was passed in 1992 in an attempt toimprove the quality of breast cancer screening with mammography and provided basicstandards that have to be met in order for a facility to be certified under the Act [19]. These included standards relevant to image quality, includingmammography machine calibration and maintenance and qualifications of mammographytechnologists. However, mammogram images are only required to be reviewed under MQSAat least once every 3 years, and MQSA inspects only a small sample of images thatare hand-picked in advance by the facility. Therefore, it could be a simple matterfor a lower-performing institution to pass image quality inspection regardless ofactual practice.

Conclusions

We found that patients of higher socioeconomic status obtained higher quality images,and higher quality images were in turn associated with earlier stage at diagnosis.In particular, results suggest that considerable gains could be made in terms ofincreasing image quality through better positioning, compression and sharpness,which could translate into earlier stage at diagnosis for patients.