Discrimination of breast cancer from healthy breast tissue using a three-component diffusion-weighted MRI model

Abstract

Purpose:

Diffusion-weighted magnetic resonance imaging (DW-MRI) is a contrast-free modality that has demonstrated ability to discriminate between pre-defined benign and malignant breast lesions. However, how well DW-MRI discriminates cancer from all other breast tissue voxels in a clinical setting is unknown. Here we explore the voxel-wise ability to distinguish cancer from healthy breast tissue using signal contributions from the newly developed three-component multi-b-value DW-MRI model.

Experimental design:

Pathology-proven breast cancer patients from two datasets (n=81 and n=25) underwent multi-b-value DW-MRI. The three-component signal contributions C₁ and C₂ and their product, C₁C₂, and signal fractions F₁, F₂ and F₁F₂ were compared to the image defined on maximum b-value (DWI_max), conventional apparent diffusion coefficient (ADC), and apparent diffusion kurtosis (K_app). The ability to discriminate between cancer and healthy breast tissue was assessed by the false positive rate given a sensitivity of 80% (FPR₈₀) and receiver operating characteristic (ROC) area under the curve (AUC).

Results:

Mean FPR₈₀ for both datasets was 0.016 (95%CI=0.008-0.024) for C₁C₂, 0.136 (95%CI=0.092-0.180) for C₁, 0.068 (95%CI=0.049-0.087) for C₂, 0.462 (95%CI=0.425-0.499) for F₁F₂, 0.832 (95%CI=0.797-0.868) for F₁, 0.176 (95%CI=0.150-0.203) for F₂, 0.159 (95%CI=0.114-0.204) for DWI_max, 0.731 (95%CI=0.692-0.770) for ADC and 0.684 (95%CI=0.660-0.709) for K_app. Mean ROC AUC for C₁C₂ was 0.984 (95%CI=0.977-0.991).

Conclusions:

The C₁C₂ parameter of the three-component model yields a clinically useful discrimination between cancer and healthy breast tissue, superior to other DW-MRI methods and obliviating pre-defining lesions. This novel DW-MRI method may serve as non-contrast alternative to standard-of-care dynamic contrast-enhanced MRI (DCE-MRI).

INTRODUCTION

Numerous studies have indicated that early breast cancer detection, with dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), has higher sensitivity than current screening programs (ultrasound and mammography) ^1–5. However, DCE-MRI has a number of limitations such as conflicting results regarding specificity ^2–6, dependency on expert radiologist readers, additional scan time and additional costs, and the use of Gadolinium-based contrast agents that are linked to deposition in the brain ⁷. In contrast, diffusion-weighted MRI (DW-MRI) does not require exogenous contrast and yields quantitative information of tissue microstructure by detecting diffusion of water molecules through application of varying degree of diffusion weighting.

Various diffusion models have demonstrated comparable ability to DCE-MRI in discriminating between pre-defined benign and malignant lesions in small regions of interest (ROIs) in the breast ^8–14. However, DW-MRI would increase its clinical utility and practicality in breast cancer screening, treatment evaluation, surgical planning, and surveillance if it could also discriminate cancer from all healthy breast tissue, not relying on lesions being pre-defined by radiologists. DW-MRI of healthy breast tissue is problematic because it consists of varying degree of admixtures of fatty and fibroglandular tissue ¹⁵ which creates an intravoxel fatty component on DW-MRI ¹⁶. Fatty tissue is primarily made up of adipocytes which contain a large lipid droplet that occupies > 90% of the cell volume, leaving only a small rim of water-containing cytoplasm. Common fat suppression techniques are designed to suppress the lipid component ¹⁷, while studies have reported very restricted diffusion in fat-suppressed healthy breast tissue ^18,19 which suggests that the water component in fatty tissue remains on conventional DW-MRI. The restricted water component in fatty tissue is especially problematic since it confounds the slow diffusion signal from intracellular cancer tissue. Thus, advanced imaging techniques are needed to discriminate cancer from all healthy breast tissue on a voxel-wise level including the restricted water component in the intravoxel fatty tissue.

Advanced, multi-component partial volume models that use extended ranges of b-values (typically up to 2000-3000 s/mm²) may theoretically isolate the slowly diffusing water pool present in cancer tissue and have become an emerging standard in several imaging domains ^20–26. Here, the DW-MRI signal is modeled as a combination of exponential decays with corresponding component apparent diffusion coefficients (ADCs), where the weighting of each component represents the attribution from a distinct pool of water from the total diffusion signal. Furthermore, selected multi-component partial volume models, such as restriction spectrum imaging (RSI) ^24–26, uses tissue-specific, pre-determined component ADCs which ensures linearization of the model, comparability across patients and rapid fitting of diffusion signal which is essential for clinical application. This is fundamentally different from conventional approaches where ADCs are not fixed but are left free and determined for each voxel independently. However, these methods are not yet well-investigated in the breast.

Initial results of multi-component partial volume models in the breast have been demonstrated by Vidić et al. ¹², showing that the normalized magnitude of the slowest component in a two-component model was excellent (AUC = 0.99) in discriminating between pre-defined benign and malignant breast lesions. Building on these findings, the multi-component model was optimized to fit the DW-MRI signal across all voxels in all breast tissue, including cancer and healthy breast tissue, resulting in a three-component model with empirical ADCs globally-determined across patients, scanners and sites ¹⁹. The three-component model was able to explain all voxels in all breast tissue, including the restricted water component in fatty tissue, rather than the averaged signal of an ROI ^27–29.

The main objective of the current study is to explore the ability of estimates derived from a three-component model to discriminate breast cancer from healthy breast tissue and to compare it to other DW-MRI methods.

MATERIALS AND METHODS

Patients

In order to validate the discriminatory power of the three-component model across scanners and sites, two datasets of pathology-proven breast cancer patients from a United States (US) site (n = 81) and a European site (n = 25) were included (Table 1). Note that 49 cases from the US site and all cases from the European site were also used to determine the three-component model with fixed ADCs for breast tissue ¹⁹. In addition, cases from the European site have been previously used for DW-MRI modeling of previously defined benign and malignant lesions ^12,30–33, linking DW-MRI signal to histological specimen ³⁴ and distortion correction techniques ³⁵. Written informed consent was obtained from patients at both sites and the studies were conducted in accordance with the Declaration of Helsinki.

Table 1.

Table of patient characteristics. ER (estrogen receptor) and PR (progesterone receptor) status were assessed by immunohistochemistry (IHC) and was considered positive if ≥1% stained nuclei was present in 10 high-power fields ⁵⁰. Human epidermal growth factor receptor 2 (HER2) status was assessed by ICH and fluorescence in situ hybridization (FISH) according to ASCO/CAP guidelines 2013 ⁵¹ or 2018 ⁵² (depending on time of recruitment); positivity was defined as an IHC score of 3+, or 2+ with a gene to chromosome ratio ≥ 2.0 by FISH. NME, non-mass enhancement.

	US dataset	European dataset
No. of patients	81	25
Median patient age, years (range)	51 (20 - 84)	53 (29 – 75)
Mean tumor volume, cm3 (range)	13.1 (0.2 - 105.9)	2.5 (0.5 – 5.8)
Histological type
Invasive carcinoma of no special type	64	17
Invasive lobular carcinoma	6	1
Tubular carcinoma	0	1
Mucinous carcinoma	0	1
Carcinoma with medullary features	0	3
Metaplastic carcinoma of no special type	4	0
Invasive papillary carcinoma	0	1
Mixed Invasive carcinoma of no special type and Invasive lobular carcinoma	3	0
Mixed Invasive carcinoma of no special type and mucinous carcinoma	1	0
Ductal carcinoma in situ	3	1
Histological grade
1	3	5
2	28	9
2/3	0	1
3	47	8
Not analyzed	3	2
ER status
Positive	53	23
Negative	27	1
Not analyzed	1	1
PR status
Positive	50	20
Negative	30	4
Not analyzed	1	1
HER2 status
Positive	14	7
Negative	64	17
Not analyzed	3	1
Lesion type mass (mass vs. NME)
Mass	67	25
NME	13	0
Mass and NME	1	0

US dataset

Ninety-five patients with pathology-proven breast cancer with no cytotoxic regimens, chemotherapy, or ipsilateral radiation therapy for this malignancy prior to MRI scanning were eligible for this prospective study. The study was approved by the Institutional Review Board of the US site. The recruitment of patients began in December 2015 and ended in June 2019. Tumor categorization was done by histopathologic analysis of core needle and open incisional biopsies. In total, 14 patients were excluded from the study; nine patients had contralateral cancer or mastectomy, one patient had no visible cancer tissue on DW-MRI, and in four patients image quality was low (low signal-to-noise ratio (n = 2), poor fat saturation (n = 1), and severe image distortion (n = 1)), resulting in 81 patients.

European dataset

This prospective study was approved by the Regional Committees for Medical and Health Research Ethics (REC Central Norway, 2011/568). The recruitment of patients began in August 2014 and ended in August 2016. Twenty-five pathology-proven breast cancer patients with inclusion criteria and tumor categorization similar to that of the US site were included; for more details, see inclusion of malignant lesions from Vidić et al. ¹².

MRI acquisition

MRI data were acquired on a 3T GE scanner (MR750, DV25-26, GE Healthcare, Milwaukee, US) and an 8-channel breast array coil with a bilateral axial imaging plane for the US dataset, while patients from the European dataset were imaged with a 3T Siemens scanner (Skyra, VD13-E11, Siemens Healthcare, Erlangen, Germany) and a 16-channel breast array coil with a unilateral sagittal imaging plane. Differences in scanner and pulse sequence parameters across sites were used to determine that the discriminatory potential of the three-component model is robust for data collected in different scanners and pulse sequence parameters. In addition to Gadolinium DCE-MRI and T2 images, both datasets included high b-value DW-MRI acquisition:

US dataset protocol:

Bilateral axial DW-MRI was performed using reduced field of view (FOV) echo-planar imaging (EPI) including the following parameters: spectral attenuated inversion recovery (SPAIR) fat suppression, TE = 82 ms, TR = 9000 ms, b-values (number of diffusion directions) = 0, 500 (6), 1500 (6), and 4000 (15) s/mm², FOV = 160 x 320 mm², acquisition matrix = 48 × 96, reconstruction matrix = 128 × 128, voxel size = 2.5 × 2.5 × 5.0 mm³, phase-encoding (PE) direction A/P, and no parallel imaging.

European dataset protocol:

Unilateral sagittal DW-MRI was performed using Stejskal-Tanner spin-echo EPI including the following parameters: FatSat (n = 15) and SPAIR (n = 10) fat suppression, TE = 88 ms, TR = 10,600 ms (n = 15) and 11,800 ms (n = 10), b-values (number of diffusion directions) = 0, 200 (6), 600 (6), 1200 (6), 1800 (6), 2400 (6), and 3000 (6) s/mm², FOV = 180 × 180 mm², acquisition matrix = 90 × 90, reconstruction matrix = 90 × 90, voxel size = 2.0 × 2.0 × 2.5 mm³, PE direction A/P, generalized auto-calibrating partially parallel acquisition (GRAPPA) with acceleration factor of 2 and 24 reference lines.

Image processing and analysis

Noise correction ³⁶ was performed to account for decreasing signal to noise ratio with increasing b-value. The observed signal (S_obs) is the mean signal across diffusion directions from one individual b-value image. Background voxels were selected by manually placing an ROI in an area in the air outside the breast on the highest b-value image, yielding the mean background intensity (S_bkg). The corrected signal intensity (S_corr) calculated from S_obs and S_bkg is given as:

$$\begin{gathered} S_{\mathit{corr}}=\sqrt{{S_{\mathit{obs}}}^2-{S_{\mathit{bkg}}}^2} \\ {S}_{\mathit{corr}}(S_{\mathit{corr}}<0)=0 \end{gathered}$$ (1)

Furthermore, corrections for eddy current artifacts, motion ²⁴ and geometric distortion ³⁶ were applied for the European dataset.

Full-volume cancer and control ROIs were manually defined on DW-MRI images, guided by all available data in the exam protocol (including DCE-MRI and anatomical T2 images, Figure 1), under supervision of and validation by two breast radiologists: RRP (US dataset) and AØ (European dataset). Cancer ROIs were drawn for the lesions corresponding to pathology-proven cancer. Control ROIs were drawn for the entire contralateral breast (US dataset) and in a cancer-free region in the ipsilateral breast at least 10 mm away from the cancer ROI (European dataset), with the aim to include all representative healthy breast tissue, excluding the axillary region, large cysts (> 2.5 cm), and susceptibility artifacts. Cancer and control ROIs were used to determine discriminatory performance between cancer and healthy breast tissue, respectively.

Figure 1.

Parameter maps for DWI_max, C₁, C₂, C₁C₂ with FPR₈₀, T2 images with cancer (red) and control (green) ROI overlay and probability density colormaps for cancer and control given C₁ and C₂ for three representative cases from the US dataset. ROIs are here only displayed for one slice but are delineated for the full volume. FPR₈₀ vary depending on the composition of healthy breast tissue in relation to the magnitude of C₁ and C₂ in cancer. (A.) Mixed tissue composition with cancer high on both dimensions. (B.) Abundant fibroglandular tissue and high C₁-magnitude of cancer. (C.) Abundant fatty tissue and high C₂-magnitude of cancer. DWI_max and C₁ performance is poorest in (C.), C₂ in (B.) while C₁C₂ has perfect performance across cases. Colormaps are given on a logarithmic scale normalized to the maximum probability density value. Y- and x-axis are defined by the maximum value for each case. Grey level windows for all images are scaled to the maximum and minimum signal intensity of each case and given in arbitrary unites. Au, arbitrary unit; C, signal contribution; DWI_max, image defined on maximum b-value; FPR₈₀, false positive rate given sensitivity of 80%; ROI, region-of-interest.

For comparison with other DW-MRI methods, the non-noise-corrected image defined on maximum b-value (DWI_max), conventional apparent diffusion coefficient (ADC), and apparent diffusion kurtosis (K_app) were estimated. DWI_max was acquired at b = 4000 s/mm² for the US dataset and b = 3000 s/mm² for the European dataset. The exponential decay formulas described by Jensen et al. ³⁸ and the corresponding b-value limits, < 1000 s/mm² and < 2000 s/mm², were used for computation of ADC and K_app maps, respectively. Note that ADC and Kapp are calculated diffusion parameters where T2 and proton density dependence are eliminated ³⁸.

To ensure that regions outside of the breast were not included in analysis, control ROIs were masked using intensity thresholding and 3D connected components (US dataset) or manually delineated within the breast boundary (European dataset) and reviewed by RRP (US dataset) and AØ (European dataset) (Figure 1 and Figure 4). Additionally, all undefined values (zero and infinite) on the image defined on b = 0 s/mm², ADC and K_app were excluded.

Figure 4.

C₁C₂, DWI_max, ADC and K_app with FPR₈₀ for discrimination between cancer (red arrowhead) and healthy breast tissue (entire cancer-free contralateral breast for the US dataset, cancer-free ipsilateral breast for the European dataset) for representative cases from the US (A.-E.) and European (F.) dataset. All cases demonstrate visual similarity between DCE-MRI and C₁C₂ maps with excellent performance compared to ADC and K_app. (A.) Excellent performance by C₁C₂ and DWI_max (B.) Excellent performance by C₁C₂ and DWI_max displaying full extent of cancer involving skin. (C.) Excellent performance by C₁C₂ and poor performance by DWI_max, ADC and K_app in a case with abundant fatty tissue. (D.) C₁C₂ improves poor DCE-MRI specificity in a case with marked background parenchymal enhancement, but partial volume artifact from the interface of fibroglandular and fatty tissue in the contralateral breast results in a low discriminatory performance. (E.) A case with NME DCIS where all diffusion maps fail; C₁C₂ has reduced cancer signal relative to the high signal from ipsilateral subarealor ducts. (F.) Sagittal image plane illustrating same trends in the European dataset. The worst FPR₈₀ for all maps is 0.9934, which would be 9,934 false positive voxels of one breast (one control ROI) approximated to contain 10,000 voxels (~30 cL). Grey level windows for all images are scaled to the maximum and minimum signal intensity of each case. ADC; conventional apparent diffusion coefficient; Au, arbitrary unit; DCE-MRI, dynamic-contrast enhanced magnetic resonance imaging; DCIS; ductal carcinoma in situ; DWImax, image defined on maximum b-value; FPR₈₀, false positive rate given sensitivity of 80%; K_app, apparent diffusion kurtosis; NME; non-mass enhancement.

Three-component modeling of diffusion signal

The corrected diffusion signal across all available b-values was fitted with a tri-exponential model, expressed as:

$$S_{corr}(b)=\text{N}[C_1\cdot e^{-b\cdot AD{C}_1}+C_2\cdot e^{-b\cdot AD{C}_2}+C_3\cdot e^{-b\cdot AD{C}_3}]$$ (2)

where S_corr is the corrected diffusion signal in arbitrary units, b is the b-value in s/mm², and C_i denotes the voxel-wise unit-less signal contribution of each component. Note that [C₁+C₂+C₃] ∝ ρ⋅exp(−TE/T2_eff), where ρ represents the proton density and T2_eff the effective T2 relaxation time in a given voxel. This model has been shown to represent the best fit across all voxels from both cancer and healthy breast tissue determined across patients, scanners and sites ¹⁹, and yielded the fixed component ADC values used in this analysis: ADC₁ = 0 mm²/s, ADC₂ = 1.4 × 10⁻³ mm²/s, and ADC₃ = 10.2 × 10⁻³ mm²/s. Fixing ADCs ensures linearization of the model and comparability of signal contributions across voxels and patients and avoids overfitting; the use of ADC₁ = 0 mm²/s means this component behaves not as a distinct exponential as in other tissue ^20,21,24,39 but as a constant offset. Hence, we use the term “three-component” for the fitted model instead of “tri-exponential”. All voxels were normalized to the 98th percentile of intensity within the b = 0 s/mm² image, indicated by the normalization factor (N). This was done to address different image intensity scaling while simultaneously preserving contribution of proton density and T2 to the DW-MRI signal.

Alternatively, the equation can be written by normalizing to the signal at b = 0 s/mm² per voxel (S(0)), yielding signal fractions (F) rather than signal contributions (C). Thus, F is related directly to diffusion and more clearly separated from proton density and T2 properties, given as;

$$S_{corr}(b)=S(0)[F_1\cdot e^{-b\cdot AD{C}_1}+F_2\cdot e^{-b\cdot AD{C}_2}+F_3\cdot e^{-b\cdot AD{C}_3}]$$ (3)

where F₁ + F₂ + F₃ = 1 and S(0) ∝ ρ⋅exp(−TE/T2_eff). This means that the signal contributions include voxel-wise T2-weighting and proton density effects, while the signal fractions are only sensitive to diffusion component effects. For the remainder of this paper, the signal at b = 0 s/mm² will be denoted S₀.

The following parametric maps were estimated from Equation 2: C₁C₂, C₁ and C₂. The parameters C₁ and C₂ were estimated directly from the model, while C₁C₂ is the corresponding product. Similarly, F₁F₂, F₁ and F₂ were estimated from Equation 3. The parametric maps C₃ and F₃ were not included due to the low cancer conspicuity of the third component ¹⁹. For completeness, the product of S₀ and signal fractions, S₀F₁F₂, S₀F₁ and S₀F₂, were estimated to investigate the relative importance of T2 and proton density effects.

Discriminating performance between cancer and healthy breast tissue

Clinical utility of the three-component derived parametric maps was assessed by comparing the voxel-wise discriminatory performance between cancer (cancer ROIs) and healthy breast tissue (control ROIs) of C₁C₂, C₁, C₂, F₁F₂, F₁, F₂, S₀F₁F₂, S₀F₁ and S₀F₂ to DWI_max, ADC, and K_app. Because there were ~52 times more healthy breast tissue voxels than cancer voxels, performance in discriminating between cancer and healthy breast tissue was examined for all voxels by the expected false positive rate given a sensitivity of 80% (FPR₈₀). In addition, the conventional performance measures receiver operating characteristic (ROC) area under the curve (AUC), sensitivity, specificity, and accuracy were estimated. Sensitivity, specificity, and accuracy were calculated for the threshold value providing optimal accuracy, defined as the mean sensitivity and specificity, assuming equal prevalence of cancer and healthy breast tissue voxels. All three-component derived parametric maps, DWI_max, and Kapp ²⁹ were assumed to have higher intensity for cancer compared to healthy breast tissue, while the opposite was assumed for ADC ^27,28. Average signal of the cancer and control ROIs were calculated, and differences were compared using a Mann-Whitney U test with a threshold significance level of 0.05.

RESULTS

Sample

The total number of voxels from cancer and healthy breast tissue from both datasets was 37,659 and 1,946,186, respectively.

Optimized three-component model parameters for discrimination

Probability density colormaps for the three-component model given C₁ and C₂ including all voxels across patients and datasets are plotted for cancer (cancer ROIs, Figure 2A) and healthy breast tissue (control ROIs, Figure 2B). These maps display two distinct probability density distributions for cancer and healthy breast tissue. The product C₁C₂ discriminates cancer from healthy breast tissue voxels, where voxels low on one or two dimensions corresponds to healthy breast tissue voxels, while cancer probability increases with increased magnitude on C₁ and C₂.

Figure 2.

Probability density colormaps for the three-component model given C₁ and C₂ including all voxels across patients and datasets are given for (A.) cancer (cancer ROIs) and (B.) healthy breast tissue (control ROIs). These maps display two distinct probability density distributions for cancer and healthy breast tissue. Cancer probability increases with increased magnitude on C₁ and C₂. Colormaps are given on a logarithmic scale normalized to the maximum probability density value. Au, arbitrary unit; C, signal contribution.

The relationship between C₁ and C₂ demonstrates that voxels with high magnitude on both dimensions had the highest probability of cancer (Figure 1A); representative cases are given in Figure 1, and all cases are given in Supplementary Figure 1–106. Discrimination performance varied depending on composition of healthy breast tissue in relation to the magnitude of C₁ and C₂ in cancer. FPR₈₀ was higher (indicating more false positive voxels) for C₁ and DWI_max in a case with abundant fat-suppressed fatty tissue and high C₂-magnitude of corresponding cancer (Figure 1C), compared to abundant fibroglandular tissue and high C₁-magnitude of corresponding cancer (Figure 1B). The opposite was seen for C₂, while C₁C₂ suppressed both fibroglandular and fatty tissue. This shows that the C₁C₂ parameter derived from the three-component model provided the optimal discrimination performance between cancer and healthy breast tissue.

All signal contributions (C₁C₂, C₁, C₂) performed better than signal fractions (F₁F₂, F₁ and F₂), given in Figure 3. Signal fractions where S₀ was included (S₀F₁F₂, S₀F₁ and S₀F₂) performed nearly equal to corresponding signal contributions (C₁C₂, C₁, C₂), see Supplementary Table 1–2.

Figure 3.

The FPR₈₀ is the false positive rate given a sensitivity of 80% for discriminating cancer and healthy breast tissue for three-component model signal contributions (C₁C₂, C₁, C₂) and signal fractions (F₁F₂, F₁, F₂), DWI_max, ADC and K_app, given per patient across the US and European dataset. Median values indicated by lines; boxes show interquartile range, block bars show data range and red crosses show outliers. The worst FPR₈₀ for all maps is 0.9934, which would be 9,934 false positive voxels of one breast (one control ROI) approximated to contain 10,000 voxels (~30 cL). ADC; conventional apparent diffusion coefficient; C, signal contribution; DWI_max, image defined on maximum b-value; F, signal fraction; K_app, apparent diffusion kurtosis.

Discriminatory performance of C₁C₂ compared to other DW-MRI methods

Mean FPR₈₀ for both datasets was 0.016 (95% CI = 0.008-0.024) for C₁C₂, 0.136 (95% CI = 0.092-0.180) for C₁, 0.068 (95 % CI = 0.049-0.087) for C₂, 0.462 (95 % CI = 0.425-0.499) for F₁F₂, 0.832 (95 % CI = 0.797-0.868) for F₁, 0.176 (95 % CI = 0.150-0.203) for F₂, 0.159 (95 % CI = 0.114-0.204) for DWI_max, 0.731 (95 % CI = 0.692-0.770) for ADC and 0.684 (95 % CI = 0.660-0.709) for K_app (Figure 3). C₁C₂ achieved the lowest FPR₈₀ with a mean ROC AUC of 0.984 (95 % CI = 0.977-0.991) when compared to other DW-MRI methods. Discriminatory performance was similar across datasets; see Supplementary Table 1–2 for all conventional performance measures and FPR₈₀ given for the two datasets separately. Average signal of the cancer and control ROIs are shown in Supplementary Table 3. All cancer and control ROIs were significantly different (p < 1 × 10⁻⁹).

C₁C₂ had excellent performance compared to ADC and K_app in a wide range of representative cases (Figure 4). DWI_max performs well in several cases (Figure 4A–B) but underperforms compared to C₁C₂, overall (Figure 3) and particularly in a case with abundant fatty tissue (Figure 4C). In addition, C₁C₂ visually improves poor DCE-MRI specificity in a case with marked background parenchymal enhancement (Figure 4D). However, C₁C₂ underperforms in cases with sparse signal from cancer, such as case of non-mass enhancement (NME) ductal carcinoma in situ (DCIS) (Figure 4E). In this case, all DW-MRI derived maps failed to identify cancer compared to healthy breast tissue. Furthermore, there was high diffusion signal from some healthy breast tissue components such as proteinaceous cysts (Figure 5B), subareolar ducts (Figure 5A), and partial volume artifact from the interface of fibroglandular and fatty tissue (Figure 4D, Figure 5C). High diffusion signal from proteinaceous cysts and subareolar ducts may be defined as nonsuspicious with the assistance of T2 images (Figure 5A–B).

Figure 5.

DCE-MRI and T2 images with corresponding C₁C₂ images illustrating false positives on C₁C₂ (yellow arrow). (A.-B.) show that false positive lesions on C₁C₂ can be defined as non-suspicious with the assistance of T2 images by a hyperintense signal on the T2 image correlated with clearly benign morphology. (A.) High signal involving subareolar ducts on T2 image and C₁C₂, not visible on DCE-MRI. (B.) Cyst visible on T2 image and C₁C₂, not visible on DCE-MRI (C.) Demonstration of limitation of C₁C₂ where background parenchymal enhancement visible on DCE-MRI and T2 image creates a partial volume artifact corresponding to the interface between fatty and fibroglandular tissue on C₁C₂. Au, arbitrary unit; C, signal contribution; DCE-MRI, dynamic-contrast enhanced magnetic resonance imaging

DISCUSSION

Our study shows that cancer can be noninvasively discriminated from healthy breast tissue using the derived parameter C₁C₂ based on a three-component DW-MRI model, with results comparable to cancer detection using DCE-MRI ^2–6 (FPR₈₀ mean = 0.016, 95 % CI = 0.008-0.024 and ROC AUC mean = 0.984, 95 % CI = 0.977-0.991). This means that C₁C₂ achieved very low false positive rates while detecting 80% or more of the defined cancer voxels. The discriminatory power of C₁C₂ was superior to that of independent signal contributions and signal fractions, conventional DW-MRI-estimates (ADC) and other methods, including diffusion kurtosis imaging (K_app) and DWI_max. The three-component model was performed across two different sites, scanners, and acquisition protocols, suggesting potential for real-world applications. The development of this advanced DW-MRI method allows for improved conspicuity of cancer relative to background breast tissue. This lays the foundation for a quantitative framework specific to pathology which may serve as an alternative to DCE-MRI.

The high discriminatory performance is due to the characteristics of the novel C₁C₂ parameter. In addition to malignancy, individual signal contributions from the three-component model were sensitive to the two primary components of healthy breast tissue: fatty (C₁) and fibroglandular (C₂) tissue. As the lipid component of fatty tissue signal is suppressed by application of fat suppression in this study (SPAIR and FatSat), we hypothesise that signal on C₁ comes from the restricted water component within adipocytes in fatty tissue ^18,19. Furthermore, neither component was sensitive to tissue with very fast diffusion properties, such as vessels, necrosis, or edema. By combining the signal contributions of the two slowest components C₁ and C₂, the majority of signal from fatty and fibroglandular tissue was suppressed so that the output image was predominantly sensitive to cancer compared to healthy breast tissue. This is particularly useful because of the varying degree of admixture of fatty and fibroglandular tissue in the breast. In fact, histological evaluation of healthy breast tissue specimen demonstrated on average 29.7 % fatty tissue component in dense breasts and 80.6 % in non-dense breasts ¹⁵. Thus, C₁C₂ may account for women with varying degree of admixed fatty tissue which is known to be an issue on conventional DW-MRI ¹⁶. While optimized for cancer discrimination, the detailed relationship between the three-component model and breast microstructure remains to be studied, as it has been for the two-component model ^34,40.

Another important aspect attributing to the high discriminatory performance is the retainment of T2 and proton density contribution to the DW-MRI signal. On conventional DW-MRI, T2 effects on DW-MRI signal is considered an inconvenience and is therefore eliminated ⁴¹. In this study we present signal contributions that include contribution from voxel-wise proton density and T2, while the signal fractions are defined to only be sensitive to diffusion effects. Thus, the importance of T2 and proton density is clearly demonstrated by the signal contributions C₁C₂, C₁ and C₂ performing far better than their signal fraction counterparts F₁F₂, F₁ and F₂. We further see these effects by signal fractions performing nearly equal to corresponding signal contributions once the signal at b = 0 s/mm², So, was included, which demonstrates that C_i ≈ S₀F_i. This has also been shown in separating benign and malignant breast lesions, where So (which has no diffusion weighting), yielded a relatively high AUC of 0.85 ¹².

We hypothesize that contributing factors to the poor performance of ADC and K_app include the restricted water component within adipocytes in fatty tissue not accounted for by fat suppression techniques and elimination of proton density and T2 effects that add to cancer discrimination. The FPR₈₀ discriminatory performance of ADC and K_app varied greatly across subjects; at best, performing around 0.2 in selected cases (Figure 4A), but overall do no better than chance. Previous studies have demonstrated significant differences between cancer and healthy breast tissue by ADC ^27,28 and Kapp ²⁹. However, these studies have been performed by signal averaged across ROIs and not voxel-wise, which does not reflect the heterogeneity of healthy breast tissue including admixture of fatty and fibroglandular tissue. Conversely, DWI_max shares the same basic properties as C₁C₂ (diffusion-, T2-, and proton density-weighting) and performs noticeably better than ADC and K_app and have several cases with perfect performance (Figure 1A–B, Figure 4A–B). However, DWI_max is also prone to influence from restricted water from fatty tissue and performs worse than C₁C₂ on average. C₁C₂ better accounts for all healthy breast tissue including the restricted water component from fatty tissue, conferring a major advantage over DWI_max and the other DWI-estimates (Figure 4C), as fibroglandular tissue is admixed with fatty tissue, and approximately 50% of women have almost entirely fatty breast tissue or scattered fibroglandular tissue ⁴².

In order for C₁C₂ to be a noninvasive alternative to DCE-MRI for breast cancer detection, it must have comparable or better sensitivity and specificity. DW-MRI is known to improve detection specificity ^8,43, which is beneficial as lesion-level DCE-MRI specificity have been reported to range from 72-97% ^2–6. In our study, performance was assessed per voxel, and the patient cohort was heterogenous, consisting of a large range of tumor volumes (mean = 10.6 cm³, range = 0.2-105.9 cm³), not reflecting the typical patient pool in the screening or surveillance setting which typically have smaller lesions. However, the high performance of discriminating cancer from all other breast tissue in comparison to other DW-MRI-based methods is highly promising and suggests clinical utility comparable to DCE-MRI. C₁C₂ may be particularly useful when DCE-MRI demonstrates false positive ⁴⁴ (Figure 4D) and false negative ⁴⁵ interpretations in patients with moderate and marked background parenchymal enhancement. Furthermore, false positive findings on C₁C₂ can be defined as nonsuspicious by a hyperintense signal on the T2 image correlated with clearly benign morphology (Figure 5A–B). While proteinaceous cysts (Figure 5B) are well known false positives on DW-MRI ⁴⁶, subareolar ducts (Figure 5A) are not commonly reported and may be due to T2 influence on C₁C₂. This indicates that C₁C₂ may assist in a non-contrast workflow with anatomical T1 and/or T2 sequences which can remove the need to administer Gadolinium contrast and any accumulation of Gadolinium in the brain ⁷.

The three-component model lays the foundation for a computationally efficient and standardized framework for breast cancer detection generalizable across sites, scanners, and acquisition protocols. By using globally-determined, fixed component ADCs, the three-component model allows for rapid fitting of diffusion signal suitable for application as a turnkey processing stream on both GE and Siemens platforms. These factors are vital for implementation in standard-of care breast MRI. Furthermore, the three-component model is performed on data acquired on extended imaging protocols (b-values up to 3000-4000 s/mm²) and requires at least three separate non-zero b-values. Inclusion of higher b-values improves discrimination by allowing better estimates of very slow diffusion characteristics of intracellular fluid within hypercellular tumors ^{9–12,20–26}. However, high b-value acquisition also results in an increased scan time, where the protocol used for the European dataset in this study for (including seven b-values up to 3000 s/mm²) had a scan time of ~8 minutes compared to a standard DW-MRI protocol (including 2 b-values) which are typically performed in 1-3 minutes. We argue that the substantially increased discriminatory performance of the derived C₁C₂ parameter compared to conventional DW-MRI justifies the increase in scan time, which is also the same scan time as conventional DCE-MRI. This does, however, illustrate the need for optimized b-value protocols for improved scan time efficiency, which is an area of interest for future development.

Several diffusion methods aim to isolate the signal from the slowly diffusing water component from cancer tissue by utilizing broad b-value ranges ^{9–12,20–26}. Diffusion kurtosis imaging is based on a simple mathematical representation of diffusion data where the derived parameter K_app has proven potential utility in the breast ^9–11,29. More advanced, multi-component partial volume models with fixed ADCs have been developed to further probe the microstructure in the brain and prostate: RSI ^24–26 (on which the three-component model is based), the vascular, extracellular, and restricted diffusion for cytometry in tumors (VERDICT) model ²¹, and the hybrid multidimensional MRI model ²². A key difference between RSI/three-component model and the hybrid multidimensional MRI model is that the hybrid model does not use pre-determined, fixed component ADCs, making comparison of corresponding signal contributions across patients and voxels difficult. Nevertheless, the hybrid model does incorporate multi-echo information not available in our study. Moreover, the T2 and proton density effects seen in RSI/three-component model are removed from the two other models, potentially reducing cancer conspicuity. Although the other multi-component partial volume models have shown promising results as cancer biomarkers in the prostate, for example, these results may be limited in breast, where fatty tissue is an important component of healthy breast tissue.

The three-component model may share biophysical similarities with the two-component intra-voxel incoherent motion (IVIM) model ⁴⁷. The two fastest component ADCs from the three-component model, ADC₂ = 1.4 × 10⁻³ mm²/s and ADC₃ = 10.2 × 10⁻³ mm²/s, are an order of magnitude apart and in the range of diffusion coefficients typically fitted for an IVIM model in breast tissue (“pure tissue diffusion coefficient” and “pseudodiffusion coefficient”) ^48,49. Therefore, we interpret that ADC₂ and the “pure tissue diffusion coefficient” from IVIM represent hindered diffusion of fibroglandular tissue, while ADC₃ and the “pseudodiffusion coefficient” from IVIM represent the very fast diffusion properties from pseudodiffusion/perfusion. This means that the optimized three-component model by Rodríguez-Soto et al. ¹⁹ is similar to an IVIM model with an additional offset C₁ with ADC₁ = 0 mm²/s which manifests in the high b-value range and accounts for the restricted water component in fatty tissue. The IVIM model focuses on perfusion properties fit to mid b-value data (typically up to 800-1000 s/mm²) and are therefore not sensitized to these very restricted diffusion properties. Moreover, as previously discussed, signal contributions include voxel-wise T2-weighting and proton density effects which is very important for discriminatory performance, while the signal fractions were only sensitive to diffusion component effects, and as such are more directly comparable to the signal fractions in an IVIM model.

There were some limitations to our study. First, the three-component methodology did not correct for partial volume artifacts which occurred at the interface between fatty and fibroglandular tissue on C₁C₂ (Figure 4D, Figure 5C). Such artifacts have the potential to be corrected, which was not investigated in this study but is an area of interest for future improvement. Another limitation concerned the definition of control ROIs; although we ensured that all control ROIs were verified as cancer-free, based on MRI review by an expert breast radiologist (both datasets) and exclusion of cases with pathology-proven contralateral cancer in the US dataset, we cannot know if occult cancer may have been included in the control ROIs. The unilateral European dataset may have been particularly prone to this, as the control ROIs were defined in the same breast as the cancer (this also made the size of control ROIs dependent on the extent of cancer and thus variable from case to case in that dataset). Lastly, detection performance is commonly evaluated at the lesion level. This study used a voxel-wise false positive rate, FPR₈₀, as its performance measure, which does not give an absolute measure comparable to other literature. However, we argue that such a measure is useful from a radiologist’s perspective, because it mimics a breast cancer examination where all voxels in the entire image are used.

In conclusion, our study is the first to demonstrate that the derived parameter C₁C₂, which is the product of the two slowest components of a three-component DW-MRI model, yields a clinically useful, noninvasive method for discrimination between cancer and healthy breast tissue. The model eliminates the need for pre-defined lesions that conventional quantitative DW-MRI metrics use and accounts for all healthy breast tissue, including the restricted water component from fatty tissue. Together with anatomical images, C₁C₂ has the potential to assist in a combined, non-contrast workflow which could serve as an alternative to DCE-MRI. The highly promising diagnostic properties were generalized across sites, scanners, and acquisition protocols, which is important for feasibility of large-scale studies for validation in routine breast cancer detection and follow-up in comparison to DCE-MRI.

TRANSLATIONAL RELEVANCE.

Here we present a novel methodology to detect cancer from surrounding healthy breast tissue. We employ an advanced diffusion-weighted magnetic resonance imaging (DW-MRI) model without the use of a contrast agent and find highly promising diagnostic properties of the derived parameter C₁C₂. The results indicate that C₁C₂ may serve as non-contrast alternative to standard-of-care dynamic contrast-enhanced MRI (DCE-MRI), which removes the need to administer Gadolinium contrast, decreasing costs and any accumulation of Gadolinium in the brain. Further clinical utility of C₁C₂ is reflected by accounting for admixed fatty tissue in healthy breast tissue and obliviation of pre-defined lesions that conventional quantitative DW-MRI metrics use. Thus, C₁C₂ may yield increased clinical utility and practicality in breast cancer evaluation, where lesions are not pre-defined. Furthermore, the diagnostic properties were generalized across sites, scanners, and acquisition protocols, which is important for feasibility of large-scale studies for validation in routine breast cancer detection.