Automatic Estimation of Volumetric Breast Density Using Artificial Neural Network-Based Calibration of Full-Field Digital Mammography: Feasibility on Japanese Women With and Without Breast Cancer

Abstract

Breast cancer is the most common invasive cancer among women and its incidence is increasing. Risk assessment is valuable and recent methods are incorporating novel biomarkers such as mammographic density. Artificial neural networks (ANN) are adaptive algorithms capable of performing pattern-to-pattern learning and are well suited for medical applications. They are potentially useful for calibrating full-field digital mammography (FFDM) for quantitative analysis. This study uses ANN modeling to estimate volumetric breast density (VBD) from FFDM on Japanese women with and without breast cancer. ANN calibration of VBD was performed using phantom data for one FFDM system. Mammograms of 46 Japanese women diagnosed with invasive carcinoma and 53 with negative findings were analyzed using ANN models learned. ANN-estimated VBD was validated against phantom data, compared intra-patient, with qualitative composition scoring, with MRI VBD, and inter-patient with classical risk factors of breast cancer as well as cancer status. Phantom validations reached an R ² of 0.993. Intra-patient validations ranged from R ² of 0.789 with VBD to 0.908 with breast volume. ANN VBD agreed well with BI-RADS scoring and MRI VBD with R ² ranging from 0.665 with VBD to 0.852 with breast volume. VBD was significantly higher in women with cancer. Associations with age, BMI, menopause, and cancer status previously reported were also confirmed. ANN modeling appears to produce reasonable measures of mammographic density validated with phantoms, with existing measures of breast density, and with classical biomarkers of breast cancer. FFDM VBD is significantly higher in Japanese women with cancer.

Keywords: Artificial neural networks (ANN); Breast tissue density; Computer analysis; Full-field digital mammography (FFDM); Image processing; Imaging phantoms; Machine learning; Magnetic resonance imaging.

Fig. 1

GEN III phantom image. Example FFDM image of GEN III phantom, with nine breast tissue-equivalent regions used in the study as references annotated with their percent volumetric breast density (VBD, also at top) and thickness combinations (also at right)

Fig. 2

An example of one GEN III phantom image’s nine tissue composition regions plotted with its fitted function for volumetric breast density (z-axis), given tissue thickness (x-axis) and signal intensity (y-axis). Each function was used to derive two separate sets of virtual phantom data using random combinations of thickness and intensity—the first for training the artificial neural network calibrations and the second for its validation

Fig. 3

Breast thickness modeling. Example breast tissue thickness map, modeled using compression paddle thickness and breast edge as described in methods. Breast tissue thickness (in z-direction, per color bar) is projected atop the bottom detector platform (0 cm, represented as black). The sum of thicknesses at all pixels makes up total breast volume (TBV)

Fig. 4

A–G Summary of mammographic breast density modeling and analysis performed in this study. Repeated imaging of a GEN III Volumetric Breast Density (VBD) phantom was performed on one FFDM system over a broad range of imaging parameters, determined by a survey of screening exams acquired as part of standard clinical practice (A). Physical phantom images were used to derive a set of virtual phantom data for the purposes of calibrating the FFDM system (B). This set of virtual calibration phantom data was used to train an artificial neural network (ANN), taking imaging parameters and data (anode and filter materials used, tube voltage, current exposure time, image exposure sensitivity, background detector signal, tissue thickness, and signal intensity) as input and outputting VBD (C). Physical phantom images were used to derive a second set of virtual phantom data for the purposes of validating the ANN calibrations of the FFDM system for calculating VBD (D). Furthermore, the original set of physical phantom images was used directly for validating the ANN calibrations (E). ANN VBD was calculated for a set of cancer and non-cancer patient’s FFDM images (F). Statistical analysis was performed fivefold: by reviewing performance on two sets of validation phantom data, intra-patient, against BI-RADS breast composition scoring, against MRI VBD, and inter-patient (G)

Fig. 5

a, b Phantom validation of ANN VBD. Linear regression results comparing ANN-estimated VBD against actual VBD (derived with virtual phantom in a, known with physical phantom in b). In considering random combinations of imaging parameters within expected ranges seen clinically, virtual phantom data derivation produced VBD values beyond 0 and 100 % VBD (a). Comparison against physical phantom data shown as a boxplot since values quantized at 0, 50, and 100 % VBD (b). Plots are inlayed with squared Pearson’s correlation coefficients and root-mean-square errors

Fig. 6

a–c Clinical images and ANN VBD maps. Example images of low- (a), mid- (b), and high-VBD (c) breasts. In each panel, the vendor post-processed image used by radiologists for diagnostic reading is shown at left (not used in study, but shown here to illustrate images used clinically and commonly used to estimate mammographic density) and the ANN-calculated VBD map is shown at right

Fig. 7

a–c Intra-patient validation of ANN VBD. Linear regression plots comparing ANN-estimated VBD measures of the left (y-axis) and right (x-axis) breast of non-cancer patients. Plots are inlayed with squared Pearson’s correlation coefficients, root-mean-square errors, and fit equations for comparisons of VBD (a), DV (b), and TBV (c)

Fig. 8

Comparison of BI-RADS breast tissue composition scoring and ANN VBD. Tukey’s multiple comparisons test reveal VBD in all categories as significantly different from each other, except between “a” and “b” as well as “a” and “c” (circles not overlapping at right, p values detailed in text)

Fig. 9

a–c Linear regression plots comparing ANN-estimated measures of VBD (y-axis) with those measured on MRI (x-axis); total breast volume (a), dense volume (b), and volumetric breast density (c). Plots are inlayed with squared Pearson’s correlation coefficients, root-mean-square errors, and fit equations for comparisons

Fig. 10

a–f Inter-patient validation of ANN VBD. Comparison plots of ANN-estimated measures of VBD with significant difference against classical risk factors of breast cancer; age (a), BMI (b, d, e), and menopause status (c). Though not reaching statistical significance, VBD compared with parity status (f) is also shown. Linear regression plots are inlayed with squared Pearson’s correlation coefficients and fit equations. Boxplots are inlayed with X ² test statistics