Dual-energy three-compartment breast imaging for compositional biomarkers to improve detection of malignant lesions.

Background: While breast imaging such as full-field digital mammography and digital breast tomosynthesis have helped to reduced breast cancer mortality, issues with low specificity exist resulting in unnecessary biopsies. The fundamental information used in diagnostic decisions are primarily based in lesion morphology. We explore a dual-energy compositional breast imaging technique known as three-compartment breast (3CB) to show how the addition of compositional information improves malignancy detection.
Methods: Women who presented with Breast Imaging-Reporting and Data System (BI-RADS) diagnostic categories 4 or 5 and who were scheduled for breast biopsies were consecutively recruited for both standard mammography and 3CB imaging. Computer-aided detection (CAD) software was used to assign a morphology-based prediction of malignancy for all biopsied lesions. Compositional signatures for all lesions were calculated using 3CB imaging and a neural network evaluated CAD predictions with composition to predict a new probability of malignancy. CAD and neural network predictions were compared to the biopsy pathology.
Results: The addition of 3CB compositional information to CAD improves malignancy predictions resulting in an area under the receiver operating characteristic curve (AUC) of 0.81 (confidence interval (CI) of 0.74-0.88) on a held-out test set, while CAD software alone achieves an AUC of 0.69 (CI 0.60-0.78). We also identify that invasive breast cancers have a unique compositional signature characterized by reduced lipid content and increased water and protein content when compared to surrounding tissues.
Conclusion: Clinically, 3CB may potentially provide increased accuracy in predicting malignancy and a feasible avenue to explore compositional breast imaging biomarkers.

Introduction

Breast cancer is the leading cause of cancer death among women globally[1]. Early detection with screening mammography has a beneficial impact on survival and has been shown to reduce cancer mortality[2-6]. However, the accuracy resulting from breast imaging technologies still has room for improvement. For instance, in the United States, 71% of biopsies do not result in a breast cancer diagnosis suggesting a modest specificity[7,8]. Furthermore, breast density affects the accuracy of full-field digital mammography (FFDM) since dense tissue can mask tumors, diminishing the sensitivity of mammography by 10–20% compared to women with fatty breasts[9]. Compared to FFDM, digital breast tomosynthesis (DBT) increases cancer detection rates and decreases recall rates. However, the added benefit of DBT is difficult to quantify and studies have demonstrated that, positive biopsy rates following screening DBT are similar to those following screening FFDM[9,10]. Also, in a registry study including over 1.5M screening mammograms from 46 registry sites, it was shown that women with the extremely dense breast tissue had neither reduced recall nor increased cancer detection rates for DBT compared to FFDM[11]. Improvements to sensitivity and specificity are needed and could result in an increase in detecting malignancies and reduction of unnecessary, benign biopsies.

The fundamental information that a radiologist uses, the attenuation of X-rays from a single exposure, has remained the same since the inception of breast imaging in 1913 (ref. [12]). Without additional information, mammography provides only relative radiopacity (i.e. tissue density relative to a background of fat) and lesion type, such as mass, asymmetry, distortion, or calcifications. Lesion classification is limited to detection of calcifications, which are often benign, as well as the shape and symmetry of high-density breast masses. Thus, lesion classification has limited reliably in predicting an invasive breast cancer. Computer-aided detection (CAD) software attempts to improve the diagnostic accuracy of mammography through the utilization of computer vision and artificial intelligence algorithms to automatically identify anomalies[13]. Yet, the fundamental information used by CAD is identical to the information radiologists use. While CAD has been shown to be clinically beneficial by some[14,15], others have shown that the addition of CAD had no significant improvement to screening sensitivity and specificity[16]. It is likely that the limit of diagnostically relevant information from radiologists and/or CAD has been reached with X-ray based, single-energy mammography, especially in women with dense breasts.

Additional diagnostic information can be obtained via contrast-enhanced mammography (CEM). Contrast imaging has demonstrated increased sensitivity to detect cancer due to differential vascularization of cancerous and benign tissue[17]. Invasive breast cancer typically presents as a mass of epithelial cells with a high degree of vascularization. IDC, and often DCIS due to its own vascularization, enhance on contrast imaging methods. However, these techniques still have low specificity because benign lesions also enhance with contrast[18-20]. Like mammography, the diagnostic information gained with contrast imaging is still based in the lesion morphology and structure of surrounding tissue. Since intravenous contrast can cause adverse effects, CEM is often used as a secondary imaging tool. Therefore, radiologists are often not afforded this information on the initial screening exam.

Radiomic features and imaging biomarkers based on tissue composition have the potential to address accuracy issues seen with current imaging techniques and technologies. Evidence suggests that the biology and atomic composition of malignant lesions differ from benign lesions and these differences manifest into macroscopically unique tissue compositions which are measurable with multispectral X-ray imaging[21,22]. First, invasive cancer is highly angiogenic and malignant tumors have been shown to consume lipids to sustain high rates of proliferation[23,24]. The central to peripheral microvasculature of the lesion differs significantly between normal tissue, fibroadenomas (FA) and different grades of invasive ductal carcinoma (IDC)[25,26]. Second, adipocytes, available at the tumor stromal interface, have demonstrated a pro-tumorigenic role for breast cancer[27]. Triple-negative cancers utilize and require fatty-acid oxidation leeched from the surrounding tissues. This has been observed using multispectral mammograms as a decrease in fat composition surrounding triple-negative cancers versus receptor-positive tumors[28]. Third, Cerussi et al.[29] found a 20% reduction in lipid, and 50% increase in water, content in invasive breast cancer versus normal breast tissue. A strong positive correlation (R = 0.98) between the macroscopic water concentration and the Scarff Bloom-Richardson Score (a histological grading scale ranging from 3 to 9 that accounts for tubule formation, nuclear pleomorphism, and mitosis counts) was also reported[30]. Fourth, invasive cancers have significantly lower X-ray attenuation than FAs that also lead to biopsy, suggesting a distinctly different composition between cancerous and benign masses[17].

The purpose of this study is to demonstrate that compositional profiles of the breast combined with CAD predictions can improve specificity of breast cancer detection. A dual-energy mammography technique known as 3-compartment breast (3CB) imaging was used to obtain the lipid–water–protein (LWP) fractions of the breast on a pixel-by-pixel basis. The 3CB scientific principals and imaging protocols have been previously presented[21,31] as well as the characteristics of malignant versus benign lesions[22,32]. To quantify the added clinical value of 3CB imaging, we compared the performance of CAD-based models to identify malignancies without and with 3CB lesion characterization. Malignant and non-malignant masses and hormone receptor status were further studied to better understand the biological mechanism which led to increased specificity of models that include 3CB composition. Compositional information from 3CB improved accuracy of malignancy predictions when compared to CAD and confirmed that invasive breast lesions have unique compositional signatures when compared to other lesion types.

Methods

Data acquisition

The participants in this study were women identified from screening and diagnostic mammography populations at the University of California San Francisco (San Francisco, CA) and H. Lee Moffitt Cancer Center and Research Institute (Tampa, FL). Women who presented with Breast Imaging-Reporting and Data System (BI-RADS) diagnostic categories 4 or 5 and who were scheduled for breast biopsies were consecutively recruited for 3CB imaging. Demographics and characteristics of this studies population are detailed in Table 1. The 3CB imaging clinical study was approved by the institutional review board at all participating research sites (University of California, San Francisco, University of Chicago, and H. Lee Moffitt Cancer Center and Research Institute) and followed Health Insurance Portability and Accountability Act-compliant protocols. All study participants provided written informed consent.

Table 1

Participant stratification by age, BMI, BI-RADS density, and duration of hormone therapy.

	N	Percentage
Participants	349	100
Age
<40	20	6
40 to <50	120	34
50 to <60	118	34
60 to <70	57	16
70 to <80	30	9
≥80	4	1
BMI
<18.5	9	3
18.5 to <25	120	34
25 to <30	92	26
≥30	128	37
BI-RADS
A	23	7
Density
B	130	37
C	162	46
D	34	10
Hormone therapy
None	321	92
<5 years	10	3
≥5 years	18	5

In addition to clinical diagnostic mammograms, participants underwent further research imaging using the 3CB protocol prior to breast biopsy. FFDMs, 2D images, were acquired on Hologic Selenia systems (Hologic, Inc., Bedford, MA). In brief, the 3CB imaging protocol consisted of two images in succession: a clinical mammogram (autocontrast, autocompression release off) and a high-energy (HE) image acquired at 39 kVp (40 mAs, 3-mm additional aluminum filtration). A calibration phantom was placed on top of the breast compression paddle during imaging to accurately estimate paddles compression depth, warp, and tilt from which exact submillimeter point thicknesses of the breast could be calculated[33]. With these three pieces of information (HE attenuation, low-energy (LE) attenuation, and local breast thickness) a system of three equations was solved which resulted in the LWP thicknesses at each pixel. Absolute accuracy of this technique has been previously verified using reference standards[21,34].

Pathology results were reported on all biopsies and radiologist delineated regions of interest (ROIs) for the mammographic abnormalities on presentation mammogram images. Participants were excluded if biopsy site annotation coordinates could not be correctly registered on presentation or 3CB images, if the lesion pathology was incomplete, or if the 3CB data set was incomplete. The 3CB protocol requires that images be acquired on calibration phantoms prior to patient imaging and the absence of calibration images or poor image quality, due to excessive movement between HE and LE image acquisition, resulted in an incomplete 3CB data set and exclusion. See Fig. 1. for a flowchart of study participant enrollment and derivation of final data set.

Fig. 1

Overview of participants and data used for modeling and analysis.

Flow diagram detailing inclusion and exclusion of data used in the final analysis. This study includes 349 patients (N) which equates to 360 biopsy sites (L) and 660 mammographic images (I) which includes craniocaudal (CC) and mediolateral oblique (MLO) views. The 660 images contained 689 radiologist delineated region of interests (ROIs) (R) and 413 computer-aided detection (CAD) delineated ROIs agreed with radiologist delineated ROIs. The final data set contained all radiologist ROIs and agreeing CAD ROIs which results in 1107 ROIs.

Table 2 stratifies ROIs by BI-RADS density categories.

Table 2

Saparation of all 689 radiologist delineated ROIs by pathology and BI-RADS density.

BI-RADS density	A	B	C	D	Total findings
Invasive	21	33	45	4	103
DCIS	2	27	22	10	61
Fibroadenoma	8	49	41	18	116
Other benign	34	164	178	33	409
Total	65	273	286	65	689

3CB feature extraction

The 3CB LWP thickness maps were generated for all FFDM images and were used to quantify the composition within the radiologist delineated ROIs. Standard presentation images and their fully registered 3CB compositional maps can be observed in Fig. 2a. Note that the 3CB images are thickness maps where each pixel corresponds to a thickness, in centimeters, of a given composition. Recall, we are investigating the diagnostic impact of independently adding compositional information to morphological features already existing in standard clinical FFDM. To abstract compositional information away from morphological features, we computationally extracted nine measurements to quantify the composition within a given region. These nine measurements included the mean, median, standard deviation, minimum, maximum, kurtosis, skew, total, and percentage value of all pixels contained within a ROI.

Fig. 2

3CB, lipid, water, and protein, data, and regions of feature extraction.

a Full presentation mammogram image and the derived three-compartment breast (3CB) thickness maps. From left to right is the standard presentation craniocaudal mammogram used for reading by a radiologist, lipid thickness map, water thickness map, and protein thickness map. Grayscale colorbars, adjacent to 3CB thickness maps, indicate thickness in cm. b The composition of the background or tissue surrounding a lesion was measured progressively by capturing three outer regions extending from the border of the lesion (yellow solid line). The outer regions extend from the lesion border at distances of 2 mm (orange dot-dashed line), 4 mm (cyan dotted line), and 6 mm (magenta dashed line). c Computer-aided detection (CAD) delineations that had some agreeance with radiologist region of interest (ROIs) (yellow line) were included in the final data set. CAD delineates suspicious masses (cyan dot-dashed line) and calcification clusters (magenta dotted line). Outer regions for all ROIs (radiologist and CAD delineated) were calculated but not displayed in this sub-figure for easy viewing.

Three additional outer ROIs were derived from the lesion ROI to capture the background or tissue immediately surrounding a lesion, see Fig. 2b. Each outer region captured all pixels extending 2 mm from the border of the previous region. Therefore, the first, second, and third outer regions contain all pixels extending from the edge of the lesion ROI out to 2 mm, the edge of the first outer region out to 2 mm, and the edge of the second outer region out to 2 mm, respectively. In other words, in relation to the lesion border, the first, second, and third outer regions measure 0–2 mm, 2–4 mm, and 4–6 mm, respectively. For each lesion, we obtained nine compositional measurements from four ROIs (lesion and three outer regions) on each of the three compositional images (3CB LWP maps) which resulted in 108 compositional features per lesion ROI.

Clinical CAD lesion detection

Low-energy, standard FFDMs were processed using commercial CAD software (SecondLook, version 7.2, iCAD, Nashua, NH) to identify suspicious masses and calcifications. The CAD software utilizes a proprietary algorithm to delineate suspicious ROIs for masses and individual calcifications as well as assigns a probability of malignancy for each delineation. Note that for input to our analysis, we used the calcification cluster ROI rather than each individual calcification ROI. Calcification cluster ROIs were calculated using the convex hull or minimum envelope which encompasses all calcifications associated with a cluster. Therefore, CAD delineated ROIs, used in our final analysis, may consist of either a suspicious mass or a calcification cluster.

Predictive modeling with morphology and 3CB

The final data set, consisting of compositional features extracted from ROIs, was split by patient ID into a train, validation, and test set using a 60, 20, and 20% split. The data were split by patient ID such that all ROIs for a given patient remained exclusively in one of the three datasets. These data split condition ensured no data leakage and ROIs from a single patient, which are highly correlated, did not end up in both the training and test set, for example. To reiterate, the train, validations, and test datasets contained their own unique subset of patients and patient ROIs and the test set contained 20% of the patients.

A neural network model was trained to predict malignancy probability from the 108 extracted 3CB features and the prediction from CAD. CAD predicts probabilities of malignancy rather than specific lesion type. To compare against CAD performance, target labels were created for our data set which combined BN and FA pathologies into a non-malignant label. ROIs with DCIS and IDC pathologies were also combined into a new malignant label. The final model was trained to output these new targets or probability of malignancy. Additional details on the neural network architecture, tuning, and hyperparameters optimization can be found in the extended Methods Section and Supplementary Fig. 1.

Quantifying the added diagnostic benefit of 3CB for malignancy prediction

The benefit of 3CB composition was evaluated using the following metrics area under the receiver operating characteristic (ROC) curve (AUC), the integrated discrimination improvement (IDI) and the net reclassification improvement (NRI)[35,36]. All metrics were computed on the unsee, independent hold out test set. The 95% confidence intervals (CI) and the mean AUC were computed via 1000 rounds of bootstrapping. All samples were selected randomly for each bootstrap round and thus the number of replacements were also random for each bootstrap round.

IDI and NRI offer additional insight into the benefits of new biomarkers beyond AUC comparison. Performance differences between a reference model and a new model, which contains the added biomarkers are evaluated across all calculated risks. The NRI measures the number of cases correctly reclassified by the new model while the IDI also takes into account the magnitude of the change in discrimination slopes. The NRI is the sum of the events NRI and the non-events NRI. In the context of this study, events and non-events correspond to malignancies and benigns, respectively. Therefore, the NRI captures the percent improvement of correctly classified malignancies and benigns by the new model which includes 3CB features. The IDI is the sum of the integrated sensitivity (IS) and the integrated 1-specificity (IP). The IS is the difference in the mean probability of malignancy for those with cancer between CAD and the neural network while the IP is the difference in the mean probability of malignancy for those with benign masses between CAD and the CAD + 3CB neural network models. The NRI and IDI changes were evaluated with respect to BI-RADS assessment categories as to investigate the clinical implications of 3CBs improvement. The BI-RADS categories of interest were 3, 4a, 4b, and 4c, with risk threshold of 2%, 10%, 50%, and 95%[37], respectively.

Lesion composition characterization

Using quantitative methods, we further investigate compositional differences among the four different lesion pathologies. To quantify these unique signatures, the median LWP values from each of the surrounding outer region ROIs were subtracted from the median LWP values from within the lesion ROI. Only radiologist drawn ROIs delineating biopsy sites were included in this analysis. Microcalcifications are present in many of the mammograms and although they are not composed of lipid, water, or protein, they can produce a high water and protein signal in the 3CB thickness maps. Therefore, the median pixel values were used to mitigate the influence microcalcifications have on the mean composition within an ROI. Lesion signatures were stratified by pathology and compositional component type (i.e. lipid, water, or protein).

Our model predicts probability of malignancy rather than lesion type, so malignant and non-malignant types were grouped for this analysis. The average signature for malignant lesion types (DCIS and IDC) and the average signature between non-malignant types (BN and FA) were computed for all outer ROIs. Differences between the malignant and non-malignant compositions were computed and p values were derived using Welch’s test for unequal variance.

We also looked at possible correlations between invasive cancers and patient hormone receptor status. It is hypothesized that cancers of different receptor type have unique compositional signatures due to utilization of exogenous fatty acids for sustained growth[28,38,39]. To investigate, we compared the composition of IDC lesions to their background and stratified each lesion by hormone receptor status. We compared triple-negative receptor lesions to all receptor-positive lesions: estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2, or any combination of the three. Differences between the triple-negative and receptor-positive lesions composition were also computed, and p values were derived using Welch’s test for unequal variance.

Table 3

Net reclassification with respects to BI-RADS risk categories.

Reference (CAD)	Events (CAD + 3CB)					Non-events (CAD + 3CB)
BI-RADS thresholds (risk range)	3 (0–≤2%)	4a (2–≤10%)	4b (10–≤50%)	4c & 5 (>50%)	Total	3 (0–≤2%)	4a (2–≤10%)	4b (10–≤50%)	4c & 5 (≥50%)	Total
3 (0–≤2%)	0	0	0	1	1	0	0	0	0	0
4a (2–≤10%)	0	0	1	0	1	2	2	8	3	13
4b (10–≤50%)	0	0	9	8	17	0	6	51	9	66
4c & 5 (≥50%)	0	0	11	20	31	0	5	47	13	65
Total	0	0	21	29	50	2	13	106	25	146

This table shows that adding 3CB allows for more accurate BI-RADS classification, as determined by probability of malignancy, for lesions with both malignant and non-malignant pathologies or events and non-events. The NRI for events and non-events is −0.02 and 0.25. The overall NRI, which is the sum of NRI events and non-events, is 0.25.

Table 4

Comparison between benign and malignant lesions.

Composition	Outer region	Malignant median	Benign median	Median difference	P value
Lipid	1	−2.50e−02	−1.82e−02	−6.82e−03	1.37e−06
Lipid	2	−4.93e−02	−3.74e−02	−1.18e−02	7.49e−08
Lipid	3	−5.56e−02	−4.03e−02	−1.53e−02	8.61e−07
Water	1	−9.39e−03	−1.77e−02	8.36e−03	6.56e−07
Water	2	−9.17e−03	−2.21e−02	1.29e−02	2.43e−07
Water	3	−5.17e−03	−2.22e−02	1.70e−02	4.17e−08
Protein	1	1.23e−02	3.44e−03	8.87e−03	1.73e−08
Protein	2	3.42e−02	1.89e−02	1.52e−02	3.66e−09
Protein	3	3.99e−02	2.33e−02	1.66e−02	7.68e−10

Difference in compositions indicated by the space between blue and orange dashed lines in Fig. 5 are quantified in this table. P values were calculated using a Welch’s test for unequal variance and all p values are significant, indicating that benign and malignant lesions have uniquely different compositions as measured by 3CB.