What do we do about mammographic density? Sarah Vinnicombe Clinical Senior Lecturer in Cancer Imaging Ninewells Hospital Medical School s.vinnicombe@dundee.ac.uk
Risk Factors for Breast Cancer Genetic factors Single autosomal dominant genetic mutations Low penetrance single nucleotide polymorphisms Environment and lifestyle factors Age at menarche, nulliparity, age at first FTP, breastfeeding, menopause, alcohol, hormones, BMI Radiation Breast density (MD)
Risk Factors for Breast Cancer Genetic factors Single autosomal dominant genetic mutations Low penetrance single nucleotide polymorphisms Environment and lifestyle Age at menarche, nulliparity, age at first FTP, breastfeeding, menopause, alcohol, hormones, BMI Radiation Breast density (MD)
Clinical Importance of MD Risk Population % Breast ca cases % x 10: BRCA 1, 2 0.5 5 x 4-10: FH++, LCIS 2-5 10 x 2-4: FH+, >50% density 20-40 50 x 0.5-2: FH, obesity, height 30-60 30 <0.5: non dense, early FTP, multiparity, breastfeeding 10-20 5 Only factor with higher relative risk at extremes is age
Objectives Can we measure MD accurately & reliably enough? Does incorporation of MD into established risk models improve their predictive ability? Is its effect independent of other known risk factors? Are there other imaging markers of risk we should consider?
Why does multimodal risk stratification matter? Marmot report 1 : 4 cases overdiagnosed for every life saved Screening should continue One size does not fit all How can screening be improved? 1 Lancet, 2012; 380:1778 86
Individualised Screening Screening plan for a given woman: Whether screening is needed When it should start Which modality Frequency of screening Age at which it should cease Lifestyle advice and interventions
How should we measure MD? Visual (BI-RADS, quintiles, Tabar, Boyd SCC, VAS) Poor reproducibility & agreement 1 Semi-automatic thresholding: area-based % MD Cumulus (research tool) Automated volumetric methods Quantra, Volpara SXA based methods 1 Lobbes et al. Insights Imaging 2012
Methods of Measurement Fully automated volumetric methods: Quantra (Hologic) Volpara (Matakina) Raw data from FFDM images Volpara: determines fat value & calculates height of fibroglandular tissue in each pixel Quantra: uses calibration model to estimate amount of fibroglandular tissue
Quantra Includes skin in estimation Yields BI-RADS equivalent score
Excludes skin Volpara
Issues: Automated Volumetric methods Skin folds (overestimation dense volume) Very dense breasts (no fat underestimation tissue volume) Very large breasts (>1 view needed) Implants Are results understandable & plausible? scoring system (eg 0-10% score 1, 11-20%, 2 etc)
Volumetric Methods Correlation with visual assessment: BI-RADS Interobserver variability for 11 readers 1 Quantra compared with dichotomised BI-RADS Best cut-off value for Quantra 22% Predicted 89% of all cases correctly Systematically lower values cf. visual classification 1 Ciatto et al. Breast 2012 epub online
Which Method should be used? RMH/Bart s Case-control study of six different methods: Cumulus, ImageJ (PD), Volpara, Quantra, SXA (FGV), BI-RADS 685, attending for screening Valid readings: 100% with Volpara, 82% with Quantra Volumetric methods: lower & smaller ranges Volpara < Quantra < SXA: medians 39ml, 71ml, 127ml Correlations driven by agreement in breast size All were associated with known breast cancer risk factors in direction expected
Do measurements depend on the digital mammography unit? S. Vinnicombe et al, ECR Vienna March 2014
Results: Visual Assessment Visual scorings No evidence of unblinding bias (mixed vs. non-mixed) 10 (10 %) of paired BI-BIRADS scores disagreed
Examples 2009 2010 2009 2010 GE Senographe DS Lorad Selenia GE Senographe DS Lorad Selenia
Correlation with risk RMH/Bart s Case-control study: 414 cases, 685 controls All methods: showed strong correlation between percent MD and risk, p for trend <0.0001 Weaker correlations with Quantra Volumetric methods measure different parameters Dos Santos Silva, Allen, Vinnicombe et al.
Correlation with risk Shepherd et al., 2011: Case-control study 275 cases, 825 controls 1 : SXA adjusted for FH, BMI, breast biopsy, age FLB OR for breast cancer risk in highest & lowest quintiles 2.5 for % density 2.9 for FGV 4.1 for % FGV FGV improved categorical risk classification in 20% Cancer Epidemiol Biomarkers Prev 2011
Conclusions Individual volumetric methods: Repeatable, reliable Correlate well with risk Agreement between methods is poor! Volpara < Quantra < SXA Same tool should be used for given woman Further work on equivalence of methods needed
Research needed Absolute measures or percent measures?? Conflicting results; absolute volume probably best 1 Does non-dense area matter? What cut-off for categorical classification? Quantra: 22% separated BI-RADS 1, 2 from 3, 4 2 Volpara: 7.5% is cut-off for BI-RADS 2 and 3 1 Keller B et al. SFBDW 2013 2 Ciatto S et al. Breast 2012
How should we use MD?
Can inclusion of MD improve risk assessment models? MD highly heritable (twin/twin studies) 1-60% of variance explained by genetic factors Only 15% of FH risk attributable to density 2 Inference: addition of MD to risk prediction tools might improve their performance 1 Boyd et al. NEJM 2002 2 Martin et al. Cancer Epidemiol Biomarkers Prev 2010
Risk Assessment Models and MD Darabi H et al. Breast Cancer Res 2012: 1,022 cases, 868 controls with complete datasets Cumulus for MD, subdivided into 6 categories Addition of MD, BMI & 18-SNP profile to Swedish Gail increased c stat from 0.569 to 0.619 (Δ AUC 0.067 highly statistically significant) Addition of MD improved AUC by 2-3%
Risk Assessment Models and MD Vachon et al., SFBDW 2013: Case-control study, 1643 cases, 2479 controls 1 77-SNP & BI-RADS density equally assoc. with risk Addition of 77-SNP score to BCSC model improved discrimination (AUC 0.69, Δ AUC 0.042) 19% of cases and 5% controls correctly reclassified
Conclusions Currently used models have only modest discriminatory ability Addition of MD to models (BI-RADS or Cumulus): only modest improvements in risk modelling Studies of addition of volumetric MD to better calibrated risk assessment models (BOADICEA, Tyrer- Cuzick) are needed
Other Markers of Risk
Other Imaging Markers of Risk DBT volumetry Texture FFDM, DBT MRI Fibroglandular volumes (FGV), water content Texture Background parenchymal enhancement Whole breast ultrasound Optical spectroscopic imaging Spectral mammography and photon counting
Alternative Modalities Digital breast tomosynthesis (DBT) more precise quantification of MD reasonable correlation with FFDM Good agreement with MR volumes, r 2 0.89 FFDM measurements are greater across BI-RADS categories 2 Tagliafico et al Breast Cancer Res Treat 2013 2 Tagliafico A et al. Br J Radiol 2013
Texture: Textural Analysis The distinctive physical composition or structure of something, especially with respect to the size, shape and arrangement of its parts Image texture: mathematically described as spatial distribution of pixel intensities Difference between maximum and minimum pixel intensity, and spacing of peaks
TA 250 200 Coarse Grey Level 150 100 50 0
Mammographic texture 3 case control studies 1,2,3 All yielded OR 2.8-14 for textural measures from SAR Independent of MD MTR marker plus computerised MD gave highest OR (5.6) with AUC 0.66 1 Only modest predictive power if added to MD model 4 (digitised FSM) 1 Nielsen M et al. Cancer Epidemiol 2010 2 Wei et al. Radiology 2011 3 Haberle et al. Breast Cancer Res 2012 4 Manduca et al. Cancer Epidemiol Biomarkers Prev 2009
DBT Texture No breast tissue overlap, better parenchymal visualisation Retroareolar ROIs, 2.5cm 3 from DBT compared with 2.5cm 2 ROI from FFDM & PD (Cumulus) 1 Skewness, coarseness, contrast, energy DBT texture better correlated with PD than FFDM texture 1 Kontos et al. Radiology 2011
MRI volumetry MRI: Volumetry % water highly correlated with MD 1 Problems: defining chest wall field inhomogeneity time-consuming gold standard?? needs automation! 1 Khazen et al Cancer Epidemiol Biomarkers Prev 2008
MRI Background parenchymal enhancement 1 1275 women, 39 cancers BPE strongly correlated with risk OR >3 (mod/marked BPE vs. min/mild BPE) Present after correction for FGV Methods of automated measurement of FGV and BPE in development 2 1 King et al. Radiology 2011 2 Wu S et al. SFBDW 2013
New: New ACR BI-RADS Lexicon fibroglandular tissue (a-d) background parenchymal enhancement (minimal-marked)
Ultrasound Whole breast US 1 Tomographic acquistion Volume-averaged sound speed (VASS) Comparator: MD on MMG VASS positively associated with MD negatively associated with non-dense area decreased with age, weight, menopausal status 1 Duric N et al. Med Phys 2013
SWE Strain produced by US probe (shear waves) Shear wave propagation captured in real time Quantifiable, good reproducibility
Optical Methods Time domain diffuse optical spectroscopy 1 Measures water, lipid, collagen, oxy- and deoxyhaemoglobin & scattering parameters Higher BI-RADS category associated with more water, collagen and less lipid More scattering with higher breast density Non-invasive method of assessing breast tissue composition 1 Taroni P et al. J Biomed Opt 2010
Research Challenges The precise biological correlate of MD Does non-dense volume matter? Are longitudinal measurements necessary? The relationship of MD, texture & spatial variation Development of automated, calibrated & validated volumetric methods - across vendors - correlated with risk - including texture measures The place of other parameters (MRI, U/S, DOSI) Incorporation into better risk assessment models
Conclusions 10 years from now Baseline assessment age 30 FH, ethnicity, reproductive history, BMI Blood/saliva test for SNP score Imaging measure of MD ABUS, DOSI, MR if high risk Computation of risk score entry into personalised screening programme
Conclusions Imaging markers of risk (MD, texture, BPE) may: Define and quantify risk But in the future will also be used to: Predict likelihood of response to adjuvant and neoadjuvant treatments Indicate chances of successful chemoprevention Inform on risk of developing recurrent/metachronous breast cancer after successful R x