Application of convolutional neural networks to breast biopsies to delineate tissue correlates of mammographic breast density.

Breast density, a breast cancer risk factor, is a radiologic feature that reflects fibroglandular tissue content relative to breast area or volume. Its histology is incompletely characterized. Here we use deep learning approaches to identify histologic correlates in radiologically-guided biopsies that may underlie breast density and distinguish cancer among women with elevated and low density. We evaluated hematoxylin and eosin (H&E)-stained digitized images from image-guided breast biopsies (n = 852 patients). Breast density was assessed as global and localized fibroglandular volume (%). A convolutional neural network characterized H&E composition. In total 37 features were extracted from the network output, describing tissue quantities and morphological structure. A random forest regression model was trained to identify correlates most predictive of fibroglandular volume (n = 588). Correlations between predicted and radiologically quantified fibroglandular volume were assessed in 264 independent patients. A second random forest classifier was trained to predict diagnosis (invasive vs. benign); performance was assessed using area under receiver-operating characteristics curves (AUC). Using extracted features, regression models predicted global (r = 0.94) and localized (r = 0.93) fibroglandular volume, with fat and non-fatty stromal content representing the strongest correlates, followed by epithelial organization rather than quantity. For predicting cancer among high and low fibroglandular volume, the classifier achieved AUCs of 0.92 and 0.84, respectively, with epithelial organizational features ranking most important. These results suggest non-fatty stroma, fat tissue quantities and epithelial region organization predict fibroglandular volume. The model holds promise for identifying histological correlates of cancer risk in patients with high and low density and warrants further evaluation.

Relationship of circulating insulin-like growth factor-I and binding proteins 1-7 with mammographic density among women undergoing image-guided diagnostic breast biopsy.

BACKGROUND: Mammographic density (MD) is a strong breast cancer risk factor that reflects fibroglandular and adipose tissue composition, but its biologic underpinnings are poorly understood. Insulin-like growth factor binding proteins (IGFBPs) are markers that may be associated with MD given their hypothesized role in breast carcinogenesis. IGFBPs sequester IGF-I, limiting its bioavailability. Prior studies have found positive associations between circulating IGF-I and the IGF-I:IGFBP-3 ratio and breast cancer risk. We evaluated the associations of IGF-I, IGFBP-3, and six other IGFBPs with MD. METHODS: Serum IGF measures were quantified in 296 women, ages 40-65, undergoing diagnostic image-guided breast biopsy. Volumetric density measures (MD-V) were assessed in pre-biopsy digital mammograms using single X-ray absorptiometry. Area density measures (MD-A) were estimated by computer-assisted thresholding software. Age, body mass index (BMI), and BMI2-adjusted linear regression models were used to examine associations of serum IGF measures with MD. Effect modification by BMI was also assessed. RESULTS: IGF-I and IGFBP-3 were not strongly associated with MD after BMI adjustment. In multivariable analyses among premenopausal women, IGFBP-2 was positively associated with both percent MD-V (ß = 1.49, p value = 0.02) and MD-A (ß = 1.55, p value = 0.05). Among postmenopausal women, positive relationships between IGFBP-2 and percent MD-V (ß = 2.04, p = 0.003) were observed; the positive associations between IGFBP-2 and percent MD-V were stronger among lean women (BMI < 25 kg/m2) (ß = 5.32, p = 0.0002; p interaction = 0.0003). CONCLUSIONS: In this comprehensive study of IGFBPs and MD, we observed a novel positive association between IGFBP-2 and MD, particularly among women with lower BMI. In concert with in vitro studies suggesting a dual role of IGFBP-2 on breast tissue, promoting cell proliferation as well as inhibiting tumorigenesis, our findings suggest that further studies assessing the role of IGFBP-2 in breast tissue composition, in addition to IGF-1 and IGFBP-3, are warranted.

Deep learning networks find unique mammographic differences in previous negative mammograms between interval and screen-detected cancers: a case-case study.

BACKGROUND: To determine if mammographic features from deep learning networks can be applied in breast cancer to identify groups at interval invasive cancer risk due to masking beyond using traditional breast density measures. METHODS: Full-field digital screening mammograms acquired in our clinics between 2006 and 2015 were reviewed. Transfer learning of a deep learning network with weights initialized from ImageNet was performed to classify mammograms that were followed by an invasive interval or screen-detected cancer within 12 months of the mammogram. Hyperparameter optimization was performed and the network was visualized through saliency maps. Prediction loss and accuracy were calculated using this deep learning network. Receiver operating characteristic (ROC) curves and area under the curve (AUC) values were generated with the outcome of interval cancer using the deep learning network and compared to predictions from conditional logistic regression with errors quantified through contingency tables. RESULTS: Pre-cancer mammograms of 182 interval and 173 screen-detected cancers were split into training/test cases at an 80/20 ratio. Using Breast Imaging-Reporting and Data System (BI-RADS) density alone, the ability to correctly classify interval cancers was moderate (AUC = 0.65). The optimized deep learning model achieved an AUC of 0.82. Contingency table analysis showed the network was correctly classifying 75.2% of the mammograms and that incorrect classifications were slightly more common for the interval cancer mammograms. Saliency maps of each cancer case found that local information could highly drive classification of cases more than global image information. CONCLUSIONS: Pre-cancerous mammograms contain imaging information beyond breast density that can be identified with deep learning networks to predict the probability of breast cancer detection.

Derived mammographic masking measures based on simulated lesions predict the risk of interval cancer after controlling for known risk factors: a case-case analysis.

PURPOSE: Women with radiographically dense or texturally complex breasts are at increased risk for interval cancer, defined as cancers diagnosed after a normal screening examination. The purpose of this study was to create masking measures and apply them to identify interval risk in a population of women who experienced either screen-detected or interval cancers after controlling for breast density. METHODS: We examined full-field digital screening mammograms acquired from 2006 to 2015. Examinations associated with 182 interval cancers were matched to 173 screen-detected cancers on age, race, exam date and time since last imaging examination. Local Image Quality Factor (IQF) values were calculated and used to create IQF maps that represented mammographic masking. We used various statistics to define global masking measures of these maps. Association of these masking measures with interval cancer vs screen-detected cancer was estimated using conditional logistic regression in a univariate and adjusted model for Breast Imaging-Reporting and Data System (BI-RADS) density. Receiver operator curves were calculated in each case to compare specificity vs sensitivity, and area under those curves were generated. Proportion of screen-detected cancer was estimated for stratifications of IQF features. RESULTS: Several masking features showed significant association with interval compared to screen-detected cancers after adjusting for BI-RADS density (up to P = 2.52E-6), and the 10th percentile of the IQF value (P = 1.72E-3) showed the strongest improvement in the area under the receiver operator curve, increasing from 0.65 using only BI-RADS density to 0.69. The highest masking group had a 32% proportion of screen-detected cancers while the low masking group had a 69% proportion. CONCLUSIONS: We conclude that computer vision methods using model observers may improve quantifying the probability of breast cancer detection beyond using breast density alone.

Longitudinal Changes in Volumetric Breast Density with Tamoxifen and Aromatase Inhibitors.

Background: Reductions in breast density with tamoxifen and aromatase inhibitors may be an intermediate marker of treatment response. We compare changes in volumetric breast density among breast cancer cases using tamoxifen or aromatase inhibitors (AI) to untreated women without breast cancer.Methods: Breast cancer cases with a digital mammogram prior to diagnosis and after initiation of tamoxifen (n = 366) or AI (n = 403) and a sample of controls (n = 2170) were identified from the Mayo Clinic Mammography Practice and San Francisco Mammography Registry. Volumetric percent density (VPD) and dense breast volume (DV) were measured using Volpara (Matakina Technology) and Quantra (Hologic) software. Linear regression estimated the effect of treatment on annualized changes in density.Results: Premenopausal women using tamoxifen experienced annualized declines in VPD of 1.17% to 1.70% compared with 0.30% to 0.56% for controls and declines in DV of 7.43 to 15.13 cm3 compared with 0.28 to 0.63 cm3 in controls, for Volpara and Quantra, respectively. The greatest reductions were observed among women with ≥10% baseline density. Postmenopausal AI users had greater declines in VPD than controls (Volpara P = 0.02; Quantra P = 0.03), and reductions were greatest among women with ≥10% baseline density. Declines in VPD among postmenopausal women using tamoxifen were only statistically greater than controls when measured with Quantra.Conclusions: Automated software can detect volumetric breast density changes among women on tamoxifen and AI.Impact: If declines in volumetric density predict breast cancer outcomes, these measures may be used as interim prognostic indicators. Cancer Epidemiol Biomarkers Prev; 26(6); 930-7. ©2017 AACR.

Relationships between mammographic density, tissue microvessel density, and breast biopsy diagnosis.

BACKGROUND: Women with high levels of mammographic density (MD) have a four- to six-fold increased risk of developing breast cancer; however, most neither have a prevalent tumor nor will they develop one. Magnetic resonance imaging (MRI) studies suggest that background parenchymal enhancement, an indicator of vascularity, is related to increased breast cancer risk. Correlations of microvessel density (MVD) in tissue, MD and biopsy diagnosis have not been defined, and we investigated these relationships among 218 women referred for biopsy. METHODS: MVD was determined by counting CD31-positive vessels in whole sections of breast biopsies in three representative areas; average MVD was transformed to approximate normality. Using digital mammograms, we quantified MD volume with single X-ray absorptiometry. We used linear regression to evaluate associations between MVD and MD adjusted for age and body mass index (BMI) overall, and stratified by biopsy diagnosis: cases (in situ or invasive cancer, n = 44) versus non-cases (non-proliferative or proliferative benign breast disease, n = 174). Logistic regression adjusted for age, BMI, and MD was used to calculate odds ratios (ORs) and 95 % confidence intervals (CIs) for associations between MVD and biopsy diagnosis. We also assessed whether the MVD-breast cancer association varied by MD. RESULTS: MVD and MD were not consistently associated. Higher MVD was significantly associated with higher odds of in situ/invasive disease (ORAdjusted = 1.69, 95 % CI = 1.17-2.44). MVD-breast cancer associations were strongest among women with greater non-dense volume. CONCLUSIONS: Increased MVD in tissues is associated with breast cancer, independently of MD, consistent with MRI findings suggestive of its possible value as a radiological cancer biomarker.

Circulating insulin-like growth factor-I, insulin-like growth factor binding protein-3 and terminal duct lobular unit involution of the breast: a cross-sectional study of women with benign breast disease.

BACKGROUND: Terminal duct lobular units (TDLUs) are the primary structures from which breast cancers and their precursors arise. Decreased age-related TDLU involution and elevated mammographic density are both correlated and independently associated with increased breast cancer risk, suggesting that these characteristics of breast parenchyma might be linked to a common factor. Given data suggesting that increased circulating levels of insulin-like growth factors (IGFs) factors are related to reduced TDLU involution and increased mammographic density, we assessed these relationships using validated quantitative methods in a cross-sectional study of women with benign breast disease. METHODS: Serum IGF-I, IGFBP-3 and IGF-I:IGFBP-3 molar ratios were measured in 228 women, ages 40-64, who underwent diagnostic breast biopsies yielding benign diagnoses at University of Vermont affiliated centers. Biopsies were assessed for three separate measures inversely related to TDLU involution: numbers of TDLUs per unit of tissue area ("TDLU count"), median TDLU diameter ("TDLU span"), and number of acini per TDLU ("acini count"). Regression models, stratified by menopausal status and adjusted for potential confounders, were used to assess the associations of TDLU count, median TDLU span and median acini count per TDLU with tertiles of circulating IGFs. Given that mammographic density is associated with both IGF levels and breast cancer risk, we also stratified these associations by mammographic density. RESULTS: Higher IGF-I levels among postmenopausal women and an elevated IGF-I:IGFBP-3 ratio among all women were associated with higher TDLU counts, a marker of decreased lobular involution (P-trend = 0.009 and <0.0001, respectively); these associations were strongest among women with elevated mammographic density (P-interaction <0.01). Circulating IGF levels were not significantly associated with TDLU span or acini count per TDLU. CONCLUSIONS: These results suggest that elevated IGF levels may define a sub-group of women with high mammographic density and limited TDLU involution, two markers that have been related to increased breast cancer risk. If confirmed in prospective studies with cancer endpoints, these data may suggest that evaluation of IGF signaling and its downstream effects may have value for risk prediction and suggest strategies for breast cancer chemoprevention through inhibition of the IGF system.

Relationship of Terminal Duct Lobular Unit Involution of the Breast with Area and Volume Mammographic Densities.

Elevated mammographic density (MD) is an established breast cancer risk factor. Reduced involution of terminal duct lobular units (TDLU), the histologic source of most breast cancers, has been associated with higher MD and breast cancer risk. We investigated relationships of TDLU involution with area and volumetric MD, measured throughout the breast and surrounding biopsy targets (perilesional). Three measures inversely related to TDLU involution (TDLU count/mm(2), median TDLU span, median acini count/TDLU) assessed in benign diagnostic biopsies from 348 women, ages 40-65, were related to MD area (quantified with thresholding software) and volume (assessed with a density phantom) by analysis of covariance, stratified by menopausal status and adjusted for confounders. Among premenopausal women, TDLU count was directly associated with percent perilesional MD (P trend = 0.03), but not with absolute dense area/volume. Greater TDLU span was associated with elevated percent dense area/volume (P trend<0.05) and absolute perilesional MD (P = 0.003). Acini count was directly associated with absolute perilesional MD (P = 0.02). Greater TDLU involution (all metrics) was associated with increased nondense area/volume (P trend ≤ 0.04). Among postmenopausal women, TDLU measures were not significantly associated with MD. Among premenopausal women, reduced TDLU involution was associated with higher area and volumetric MD, particularly in perilesional parenchyma. Data indicating that TDLU involution and MD are correlated markers of breast cancer risk suggest that associations of MD with breast cancer may partly reflect amounts of at-risk epithelium. If confirmed, these results could suggest a prevention paradigm based on enhancing TDLU involution and monitoring efficacy by assessing MD reduction.

Leukocyte telomere length and its association with mammographic density and proliferative diagnosis among women undergoing diagnostic image-guided breast biopsy.

BACKGROUND: Elevated mammographic density (MD) is a strong breast cancer risk factor but the mechanisms underlying the association are poorly understood. High MD and breast cancer risk may reflect cumulative exposures to factors that promote epithelial cell division. One marker of cellular replicative history is telomere length, but its association with MD is unknown. We investigated the relation of telomere length, a marker of cellular replicative history, with MD and biopsy diagnosis. METHODS: One hundred and ninety-five women, ages 40-65, were clinically referred for image-guided breast biopsies at an academic facility in Vermont. Relative peripheral blood leukocyte telomere length (LTL) was measured using quantitative polymerase chain reaction. MD volume was quantified in cranio-caudal views of the breast contralateral to the primary diagnosis in digital mammograms using a breast density phantom, while MD area (cm(2)) was measured using thresholding software. Associations between log-transformed LTL and continuous MD measurements (volume and area) were evaluated using linear regression models adjusted for age and body mass index. Analyses were stratified by biopsy diagnosis: proliferative (hyperplasia, in-situ or invasive carcinoma) or non-proliferative (benign or other non-proliferative benign diagnoses). RESULTS: Mean relative LTL in women with proliferative disease (n = 141) was 1.6 (SD = 0.9) vs. 1.2 (SD = 0.6) in those with non-proliferative diagnoses (n = 54) (P = 0.002). Mean percent MD volume did not differ by diagnosis (P = 0.69). LTL was not associated with MD in women with proliferative (P = 0.89) or non-proliferative (P = 0.48) diagnoses. However, LTL was associated with a significant increased risk of proliferative diagnosis (adjusted OR = 2.46, 95% CI: 1.47, 4.42). CONCLUSIONS: Our analysis of LTL did not find an association with MD. However, our findings suggest that LTL may be a marker of risk for proliferative pathology among women referred for biopsy based on breast imaging.

The effect of change in body mass index on volumetric measures of mammographic density.

BACKGROUND: Understanding how changes in body mass index (BMI) relate to changes in mammographic density is necessary to evaluate adjustment for BMI gain/loss in studies of change in density and breast cancer risk. Increase in BMI has been associated with a decrease in percent density, but the effect on change in absolute dense area or volume is unclear. METHODS: We examined the association between change in BMI and change in volumetric breast density among 24,556 women in the San Francisco Mammography Registry from 2007 to 2013. Height and weight were self-reported at the time of mammography. Breast density was assessed using single x-ray absorptiometry measurements. Cross-sectional and longitudinal associations between BMI and dense volume (DV), non-dense volume (NDV), and percent dense volume (PDV) were assessed using multivariable linear regression models, adjusted for demographics, risk factors, and reproductive history. RESULTS: In cross-sectional analysis, BMI was positively associated with DV [ß, 2.95 cm(3); 95% confidence interval (CI), 2.69-3.21] and inversely associated with PDV (ß, -2.03%; 95% CI, -2.09, -1.98). In contrast, increasing BMI was longitudinally associated with a decrease in both DV (ß, -1.01 cm(3); 95% CI, -1.59, -0.42) and PDV (ß, -1.17%; 95% CI, -1.31, -1.04). These findings were consistent for both pre- and postmenopausal women. CONCLUSION: Our findings support an inverse association between change in BMI and change in PDV. The association between increasing BMI and decreasing DV requires confirmation. IMPACT: Longitudinal studies of PDV and breast cancer risk, or those using PDV as an indicator of breast cancer risk, should evaluate adjustment for change in BMI.

Relationship of serum estrogens and metabolites with area and volume mammographic densities.

Elevated mammographic density is a breast cancer risk factor, which has a suggestive, but unproven, relationship with increased exposure to sex steroid hormones. We examined associations of serum estrogens and estrogen metabolites with area and novel volume mammographic density measures among 187 women, ages 40-65, undergoing diagnostic breast biopsies at an academic facility in Vermont. Serum parent estrogens, estrone and estradiol, and their 2-, 4-, and 16-hydroxylated metabolites were measured using liquid chromatography-tandem mass spectrometry. Area mammographic density was measured in the breast contralateral to the biopsy using thresholding software; volume mammographic density was quantified using a density phantom. Linear regression was used to estimate associations of estrogens with mammographic densities, adjusted for age and body mass index, and stratified by menopausal status and menstrual cycle phase. Weak, positive associations between estrogens, estrogen metabolites, and mammographic density were observed, primarily among postmenopausal women. Among premenopausal luteal phase women, the 16-pathway metabolite estriol was associated with percent area (p = 0.04) and volume (p = 0.05) mammographic densities and absolute area (p = 0.02) and volume (p = 0.05) densities. Among postmenopausal women, levels of total estrogens, the sum of parent estrogens, and 2-, 4- and 16-hydroxylation pathway metabolites were positively associated with area density measures (percent: p = 0.03, p = 0.04, p = 0.01, p = 0.02, p = 0.07; absolute: p = 0.02, p = 0.02, p = 0.01, p = 0.02, p = 0.03, respectively) but not volume density measures. Our data suggest that serum estrogen profiles are weak determinants of mammographic density and that analysis of different density metrics may provide complementary information about relationships of estrogen exposure to breast tissue composition.

Comparison of mammographic density assessed as volumes and areas among women undergoing diagnostic image-guided breast biopsy.

BACKGROUND: Mammographic density (MD), the area of non-fatty-appearing tissue divided by total breast area, is a strong breast cancer risk factor. Most MD analyses have used visual categorizations or computer-assisted quantification, which ignore breast thickness. We explored MD volume and area, using a volumetric approach previously validated as predictive of breast cancer risk, in relation to risk factors among women undergoing breast biopsy. METHODS: Among 413 primarily white women, ages 40 to 65 years, undergoing diagnostic breast biopsies between 2007 and 2010 at an academic facility in Vermont, MD volume (cm(3)) was quantified in craniocaudal views of the breast contralateral to the biopsy target using a density phantom, whereas MD area (cm(2)) was measured on the same digital mammograms using thresholding software. Risk factor associations with continuous MD measurements were evaluated using linear regression. RESULTS: Percent MD volume and area were correlated (r = 0.81) and strongly and inversely associated with age, body mass index (BMI), and menopause. Both measures were inversely associated with smoking and positively associated with breast biopsy history. Absolute MD measures were correlated (r = 0.46) and inversely related to age and menopause. Whereas absolute dense area was inversely associated with BMI, absolute dense volume was positively associated. CONCLUSIONS: Volume and area MD measures exhibit some overlap in risk factor associations, but divergence as well, particularly for BMI. IMPACT: Findings suggest that volume and area density measures differ in subsets of women; notably, among obese women, absolute density was higher with volumetric methods, suggesting that breast cancer risk assessments may vary for these techniques.

Interpolation-based super-resolution reconstruction: effects of slice thickness.

Standard clinical magnetic resonance imaging (MRI) is acquired in two-dimensions where the in-plane resolution is higher than the slice select direction. These acquisitions include axial, coronal, and sagittal planes. To date, there have been few attempts to combine the information of these three orthogonal orientations. This paper aims to take advantage of the different in-plane resolution acquired from each plane orientation and combine them into one volume in order to attain a higher resolution image. This combination of MRI data will allow the detection of smaller areas that would otherwise be missed using only one slice orientation. A comparison of slice thicknesses along with image registration is performed. The mean-squared error and peak signal-to-noise were computed for quantitative assessment. MRI and phantom scans and joint histograms were used for qualitative assessment.

Evaluation of interpolation effects on upsampling and accuracy of cost functions-based optimized automatic image registration.

Interpolation has become a default operation in image processing and medical imaging and is one of the important factors in the success of an intensity-based registration method. Interpolation is needed if the fractional unit of motion is not matched and located on the high resolution (HR) grid. The purpose of this work is to present a systematic evaluation of eight standard interpolation techniques (trilinear, nearest neighbor, cubic Lagrangian, quintic Lagrangian, hepatic Lagrangian, windowed Sinc, B-spline 3rd order, and B-spline 4th order) and to compare the effect of cost functions (least squares (LS), normalized mutual information (NMI), normalized cross correlation (NCC), and correlation ratio (CR)) for optimized automatic image registration (OAIR) on 3D spoiled gradient recalled (SPGR) magnetic resonance images (MRI) of the brain acquired using a 3T GE MR scanner. Subsampling was performed in the axial, sagittal, and coronal directions to emulate three low resolution datasets. Afterwards, the low resolution datasets were upsampled using different interpolation methods, and they were then compared to the high resolution data. The mean squared error, peak signal to noise, joint entropy, and cost functions were computed for quantitative assessment of the method. Magnetic resonance image scans and joint histogram were used for qualitative assessment of the method.