Evaluation of a new image reconstruction method for digital breast tomosynthesis: effects on the visibility of breast lesions and breast density

Abstract

Objective:

To compare image quality and breast density of two reconstruction methods, the widely-used filtered-back projection (FBP) reconstruction and the iterative heuristic Bayesian inference reconstruction (Bayesian inference reconstruction plus the method of total variation applied, HBI).

Methods:

Thirty-two clinical DBT data sets with malignant and benign findings, n = 27 and 17, respectively, were reconstructed using FBP and HBI. Three experienced radiologists evaluated the images independently using a 5-point visual grading scale and classified breast density according to the American College of Radiology Breast Imaging-Reporting And Data System Atlas, fifth edition. Image quality metrics included lesion conspicuity, clarity of lesion borders and spicules, noise level, artifacts surrounding the lesion, visibility of parenchyma and breast density.

Results:

For masses, the image quality of HBI reconstructions was superior to that of FBP in terms of conspicuity,clarity of lesion borders and spicules (p < 0.01). HBI and FBP were not significantly different in calcification conspicuity. Overall, HBI reduced noise and supressed artifacts surrounding the lesions better (p < 0.01). The visibility of fibroglandular parenchyma increased using the HBI method (p < 0.01). On average, five cases per radiologist were downgraded from BI-RADS breast density category C/D to A/B.

Conclusion:

HBI significantly improves lesion visibility compared to FBP. HBI-visibility of breast parenchyma increased, leading to a lower breast density rating. Applying the HBIR algorithm should improve the diagnostic performance of DBT and decrease the need for additional imaging in patients with dense breasts.

Advances in knowledge:

Iterative heuristic Bayesian inference (HBI) image reconstruction substantially improves the image quality of breast tomosynthesis leading to a better visibility of breast carcinomas and reduction of the perceived breast density compared to the widely-used filtered-back projection (FPB) reconstruction. Applying HBI should improve the accuracy of breast tomosynthesis and reduce the number of unnecessary breast biopsies. It may also reduce the radiation dose for the patients, which is especially important in the screening context.

Introduction

Digital breast tomosynthesis (DBT) has been shown to overcome some limitations of standard two-dimensional full-field digital mammography (FFDM) that are caused by the overlap of normal and pathological tissues.^1–7 Several studies have demonstrated the advantages of DBT for breast cancer screening, such as increased cancer detection rates and reduces callback rates.^8,9 It has also been hypothesized that breast density, which is an image biomarker of tissue composition and an independent risk factor for breast cancer, can be more accurately determined by DBT.^10,11 In DBT, volumetric reconstruction of the breast is typically obtained from a finite number of low-dose projections at different X-ray tube angles. Using a wider scan range for DBT has shown to have a positive effect on image quality reducing artifacts and increasing depth resolution.¹² It should further improve the detection of breast masses by reducing superimposed breast anatomy.^1,4 However, it may be associated with an elevated radiation doseage to the breast.¹³ The acquired projection data are usually reconstructed as 1 mm slices using either filtered-back projection (FBP) or iterative reconstruction algorithms. At present, the diagnostic accuracy of DBT has almost exclusively been determined with systems incorporating FBP algorithms¹⁴ , despite the fact that iterative reconstruction (IR) have played key roles in other fields of three-dimensional medical imaging, such as computed tomography. Studies using IR confirmed a better detectability of pathologies and reduction of the patient’s radiation dose by suppressing image noise.^15,16 IR improves image quality through cyclic image processing and has the advantage of allowing physical effects to be modelled, accounting for the probability distribution of the experimental measurements. Although all available IR solutions generally speaking reduce artifacts and radiation dose, the magnitude of these effects depends on the specific IR algorithm. IR techniques could potentially be particularly useful for mammography. Since various artifacts can mimic or obscure pathological changes and reduce the sensitivity or specificity of a modality,¹⁷ IR can identify subtle pathological changes through variations in tissue attenuation properties. To date, pure iterative algorithms are rarely used in a clinical DBT setting and are usually only applied in DBT systems employing sparse sampling, despite the fact that phantom-based studies have demonstrated promising results.^12,18,19

Recently, FBP in DBT was compared using two IR algorithms called maximum likelihood expectation (MLEM) and simultaneous iterative reconstruction technique (SIRT). MLEM has shown to provide a good balance between visibility of high-frequency components (like calcifications) and low-frequency components (like soft-tissue lesions).²⁰ Another study examined a variant of iterative total variation minimization (TVM) reconstruction, where it was shown that this technique preserved lesion contrast and high image quality while at the same time reducing the number of projections needed (and hence reducing radiation dose to the breast).²¹ In a third study, a Compton scattering suppression-based method for DBT based on Bayesian estimations was found to be superior to both SIRT and FBP in terms of object contrast.²²

In this study, we evaluate an IR method called heuristic Bayesian inference (HBI) reconstruction, which combines bayesian estimates and the TVM algorithm.^21,22 We compare image reconstruction by HBI and FBP in a series of patient cases in terms of breast lesion characteristics and breast density assessment.

Methods and materials

Study population

This retrospective study was approved by the National Federal Radiation Commission and the Institutional Review Board, Germany. Data were analysed in accordance with the Health Insurance Portability and Accountability Act and the Declaration of Helsinki. Patient cases were retrospectively selected from the Institute of Clinical Radiology and Nuclear Medicine, University Medical Center Mannheim, Heidelberg University, Germany. The study population included 14 females examined during a 17 month period from February 2016 to July 2017. All patients had suspicious findings on full-field digital mammography (BI-RADS category 4 and 5) and DBT was performed for further diagnostic workup. 3/14 (21.4.%) patients underwent mammography for routine screening, 4/14 (28.6%) for aftercare following contralateral breast cancer and 10/14 (71.4%) patients presented with suspicious clinical symptoms and mammography was obtained for further diagnostic workup.

In 12 patients DBT was performed unilaterally. Two females had suspect findings in both breasts and DBT was obtained bilaterally. Subsequently, 16 breasts were scanned. Two projection views (medio-lateral oblique and craniocaudal) were aquiered per breast. Finally, 32 consecutive data sets were included in the evaluation.

Histopathology served as reference standard for all primary lesions that led to the categorization BI-RADS 4 or 5. In case of additional lesions, that appear most likely benign on mammography a combination of breast MRI, ultrasound and mammography as well as follow-up imaging of at least 2 years were used to verify benign findings. Further, histopathology of benign findings was available for four breasts, since patients underwent total mastectomy for cancer cure.

Overall, 44 radiographic findings were registered and evaluated of which 27 were cancers and 17 were benign. Radiation dose data were extracted from the DICOM headers of the images.

Image acquisition

DBT projection data of 32 DBT scans of 16 breasts were extracted from the picture archiving and communication system at the Institute of Clinical Radiology and Nuclear Medicine, University Medical Center Mannheim (Heidelberg University, Germany). DBT images were aquired in mediolateral-oblique (MLO) and the craniocaudal (CC) projection views using the wide-angle DBT device Mammomat Inspiration by Siemens (Siemens Healthcare Sector, Erlangen, Germany). In each projection view, 25 projection images were acquired over an angular range of approximately 50° (±25° around the MLO/CC position) using an anode/filter combination of W/Rh while the detector was stationary.

Reconstruction methods applied to patient data

The 25 unprocessed projection images were used for tomographic reconstructions. Graphics processing unit (GPU)-based reconstruction was performed using a NVIDIA GeForce GTX 470 (1280 MB) graphics card installed in an ACPI Multiprocessor x64-based PC [Microsoft Windows XP Professional x64 Edition operating system, DualCore Intel Pentium D 945 processor, 3417 MHz, system Board: MSI 975X Platinum PowerUp Edition (MS-7246), system memory: 7296 MB].

Filtered-back projection reconstruction

For the FBP reconstructions, each projection image was filtered with a ramp filter (i.e., von Hann filter, spectral filter) to suppress high frequencies and a slice thickness filter to smooth the images. Images were then back-projected using cone beam geometry. This FBP method was an optimized variant integrated on commercially available Siemens Inspiration units.²³

Heuristic Bayesian Inference reconstruction

The statistical iterative method we tested, HBI reconstruction, is similar to a previously described method²⁴ but differs in its correction to minimise the residuals of the radial integrals while maintaining the intensities (i.e. by using the method of total variation minimization).²¹ The HBI reconstruction algorithm can be written as:

$${\displaystyle x_j^{\left(k+1\right)}=x_j^{\left(k\right)}\left[1+{\lambda }^{\left(k\right)}\sum _{i=1}^I\frac{A_{ij}}{\sum _{j=1}^J{A}_{ij}}\frac{⟮p_i-\sum _{j=1}^J{A}_{ij}{x}_j^{(k)}⟯}{\sum _{j=1}^J{A}_{ij}{x}_j^{(k)}}\right],}$$

where ${\displaystyle x_j^{(k)}}$ is the value of the j-th component of the vector of unknowns at the k-th iteration,${\displaystyle A_{ij}}$ is elements of the projection matrix, ${\displaystyle p_i}$ is the value of the measured radial integral in the pixel number i, and ${\displaystyle {\lambda }^{(k)}}$ is sequence relaxation parameter values (0 < ${\displaystyle {\lambda }^{(k)}}$< 1). The parameter${\displaystyle {\lambda }^{(k)}}$ for k = 0 was 0.3, while for k = 1 it was 0.15. In prior optimization work,²⁵ two iterations was found to yield the best anatomic reproduction and was therefore used in the current work.

Analysis of reconstruction methods

Three expert radiologists with 10, 2 and 6 years of experience in breast imaging interpretation independently evaluated the images on a clinical workstation using two five megapixel monitors that had been routinely calibrated yearly. The reviewing radiologists were blinded to any lesion-specific information, to the reconstruction method used and to the order the images were displayed. The radiologists were free to alter the window and level settings and use the zoom and pan function. No restriction was set on the interpretation time.

The analysis included the following steps:

Review of lesion-specific features

Each breast lesion was evaluated in a side-by-side-review using a 5-point scale (1 = much worse, 2 = slightly worse, 3 = equal, 4 = slightly better, 5 = much better) for the following attributes:

• clarity of lesion edge/spiculations

• noise level (distracting quantum-like noise surrounding the lesion)

• artifact suppression (how severe the artifacts around the lesion were)

• lesion conspicuity (how well the lesion contrasted with neighboring tissues)

• visibility of fibroglandular parenchyma (how much it stood out from fatty tissues)

Breast density assessment

Breast density was classified according to the American College of Radiology Breast Imaging-Reporting And Data System Atlas^® fifth edition,²⁶ which has the following categories: A = The breasts are almost entirely fatty; B = There are scattered areas of fibroglandular density; C = The breasts are heterogeneously dense (which may obscure small masses); D = The breasts are extremely dense (which lowers the sensitivity of mammography). The breast density assessment was performed in a blinded setting at least 1 week after the side-by-side-review, and one single DBT case was interpreted at a time. The cases were displayed and interpreted in random order in two reading sessions. A minimum of one and a half week was allowed between the reading sessions to eliminate potential memory effects.²⁷

Statistical analysis

Statistical analyses of the radiologists’ interpretations were done using the visual grading characteristics (VGC) analyzer software.²⁸ An area under the visual grading curve (AUC_VGC) was computed to measure image quality of the two image types being compared with 95% confidence intervals generated using bootstrapping.²⁹ AUC_VGC ranged from 0 to 1. An AUC_VGC value of 0.5 reflected similar image quality for the two image reconstruction types, while an AUC_VGC > 0.5 indicated the HBI-image quality to be higher than the conventional FBP method, and an AUC_VGC < 0.5 indicated the HBI image was lower quality. Image quality was considered significantly different from the reference settings when the 95% confidence intervals (within brackets) did not enclose the dashed line (AUC_VGC = 0.5), which is analogous to a p-value that is below an α of 0.05. Analyses was paired in terms of observations made on lesions, breast parenchyma, and breast density. Stratified analysis was performed by radiographic pattern and histopathology. Radiologists’ inter-reader agreement of breast density assessments were analyzed using Fleiss’ κ statistics.³⁰

Results

Study population

The study population characteristics and radiographic pattern of each lesion type are presented in Table 1. All findings were visible on both reconstruction methods and were hence included in the side-by-side-analysis.²⁸ The mean pathological tumor size was 19.7 mm (4–75). Average glandular dose (AGD) was 0.72 mGy (0.4–1.22 mGy) per DBT view and 1.46 mGy (0.81–2.42) per breast. The median patient age was 62 (range 48–80 years).

Table 1. .

Study population characteristics and radiographic presentation of histopathological lesion types.

Parameter		Radiographic presentation
		Spiculated mass	Indistinct mass	Well-circumscribed mass	Architectural distortion	Calcifications
Lesion type	n (%)
Carcinoma of no special type	19	11	7		1
Invasive lobular carcinoma	4	1	3
Invasive tubular carcinoma	1	1
Ductal carcinoma in situ	3					3
Benign calcifications	8					8
Fibroadenoma	2			2
Cysts	7			7
Total	44	12	10	10	1	11

Image quality of radiographic findings

Image quality (AUC_VGC) is presented for masses and calcifications (Table 2). For masses, HBI was rated significantly superior to FBP with an AUC_VGC > 0.5 for noise level, extent of artifacts, clarity of lesion edge and lesion conspicuity. For calcifications, HBI resulted in lower noise surrounding the lesion and a higher suppression of in-plane artifacts. Details of calcifications (conspicuity and border characteristics) were comparable for the two reconstruction methods (AUC = 0.5; p > 0.05). The visibility of fibroglandular parenchyma was significantly higher for the HBI method than the FBP method (AUC_VGC = 0.82, p < 0.01).

Table 2. .

Image quality assessment of the radiographic findings. Image quality metric is the AUC_VGC, which is presented with 95% CIs. When the CI values do not enclose 0.5, the difference is statistically significant.

	Radiographic findings
	Masses (n = 33)				Calcifications (n = 11)
	Reader #				Reader #
Image quality Parameter	1	2	3	All	1	2	3	All
Lesion conspicuity	0.53 [0.42–0.64]	0.61 [0.52–0.68]	0.74 [0.63–0.83]	0.63 [0.57–0.68]	0.47 [0.27–0.67]	0.50 [0.40–0.60]	0.57 [0.50–0.67]	0.51 [0.42–0.60]
Clarity of lesion edges	0.83 [0.74–0.90]	0.66 [0.53–0.77]	0.73 [0.62–0.83]	0.74 [0.67–0.81]	0.46 [0.23–0.70]	0.43 [0.30–0.60]	0.57 [0.50–0.67]	0.49 [0.40–0.60]
Noise level	0.96 [0.91–0.99]	0.85 [0.75–0.93]	0.89 [0.80–0.97]	0.90 [0.85–0.95]	0.67 [0.43–0.87]	0.80 [0.60–0.93]	0.57 [0.50–0.67]	0.68 [0.57–0.79]
Artifact suppression	0.91 [0.86–0.97]	0.78 [0.68–0.87]	0.92 [0.87–0.97]	0.87 [0.83–0.91]	0.68 [0.38–0.94]	0.67 [0.54–0.81]	0.68 [0.54–0.81]	0.67 [0.54–0.78]

AUC_VGC, area under the visual grading curve; CI, confidence interval.

Subanalysis of lesion-specific image quality by radiographic pattern and histopathology

Figure 1 presents the image quality parameters (AUC_VGC) correlated to (a) the specific radiographic pattern of each lesion and (b) the histopathology of each lesion. HBI significantly improved the image quality of all masses regardless of their morphologic subtype. This was also the case for the lesions’ borders, image noise level and severity of artifacts surrounding the lesion. The lesion’s conspicuity was increased for HBI relative to FBP for all readers, but this increase was not statistically significant. For lesion histopathology, incremental effects of HBI reconstructions were found for malignant as well as benign lesions. Again, the lesions’ borders, image noise level and artifact supression improved for all histopathological subtypes except for ductal carcinomas in situ, where there was no difference in the lesions border visibility.

Figure 1. .

Subanalysis of the two reconstruction methods according to (a) radiographic pattern and (b) microscopic lesion type. A performance indicator (AUC_VGC) larger than 0.5 indicate a higher quality for the iterative Bayesian method when compared with the standard filtered-back projection method. Where the 95% confidence intervals do not enclose the line at 0.5, the difference is statistically significant. AUC_VGC,area under the visual grading curve; DCIS, ductal carcinomas in situ; ILC, invasive lobular carcinoma.

There was only one case of a carcinoma of no special type that presented as architectural distortion (the statistical power is low and the 95% confidence intervals wide). Nevertheless, even for this case, HBI images provided improved characteristics of the lesions’ borders, the image noise level and the severity of artifacts surrounding the lesion, but with no significant effect on conspicuity.

Figures 2–6 illustrate patients with breast lesions and the different types of DBT image reconstruction methods.

Figure 2. .

A 50-year-old patient with two spiculated masses in the upper and lower quadrant of her left breast. Histology confirmed a bifocal carcinoma of no specific type (12 mm and 25 mm; indicated by white arrows). (a). Image reconstructed using FBP. (b). Image reconstructed using the iterative Bayesian-based method (HBI). Visibility of the fibroglandular tissue was rated higher on the HBI method by all three radiologists.

Figure 3. .

A unifocal 29 mm invasive lobular carcinoma. (a) FBP reconstruction. (b) HBI reconstruction. The noise level and the artifacts surrounding the lesion and the image noise were significantly reduced using the HBI method.

Figure 4. .

Magnification view of the right breast containing a single macrocalcification typical for a calcified fat necrosis and a well-circumscribed oval mass (a cyst). (a) FBP reconstruction. (b) HBIR reconstruction. Artifacts (dark signals indicated with white arrows) surrounding the lesions in scan direction and skinline (brighter signals indicated with an arrow) are less pronounced with the HBIR method.

Figure 5. .

Magnification view of finer microcalcifications (from the left to the right; indicated with an arrow):~200 and 425 microns within a 22 mm invasive ductal carcinoma using (a) FBP and (b) HBI.

Figure 6. .

Well-circumscribed lobulated lesion. (a) FBP reconstruction. (b) HBI reconstruction. The overall noise level was significantly more pronounced when using the FBP method, which can affect the lesion edge (see also Table 2). Breast MRI confirmed a 5 mm cyst.

Breast density assessment

Assessed breast density categories were significantly lower using HBI compared to FBP (p = 0.016). On average, five cases per radiologist were downgraded when HBI was used, three of which changed from dense breasts (breast density category C or D) to fatty breasts (breast density category category A or B) (Figure 7). Figure 8 illustrates a case where the breast density category was downgraded using the HBI reconstruction. On average, two cases per radiologist were upgraded when HBI was used. Readers of FBP and HBI reconstructed images had a fair level of agreement (K_FBP = 0.375 and K_HBI = 0.356).

Figure 7. .

Breast density BI-RADS categorized using either (a) FPB or (b) HBI reconstructed digital breast tomosynthesis images for the three readers.

Figure 8. .

DBT MLO view of the left breast. (a) FBP reconstruction. (b) HBI reconstruction. Breast density was categorized as ACR BI-RADS C by all radiologists reviewing the FBP-reconstructions, which was downgraded to ACR BI-RADS B by two of the radiologists when reviewing the HBIR-reconstructions, likely due to a clearer reproduction of fatty and fibroglandular tissues.

Discussion

In this study, the HBI reconstruction and concurrent FBP reconstruction were compared for 32 DBT data sets aquired with a wide-angle DBT system. To the authors’ knowledge this study is the first to evaluate these algorithms, which were applied on DBT data sets that were obtained in a clinical setting for the diagnostic workup of patients with unclear findings on conventional mammography.

We found that the HBI method significantly improved the image quality of breast lesions in DBT, which is important because increased image quality could in turn improve diagnostic accuracy. For all lesion types, two parameters in particular were significantly improved, namely the image noise level and the artifacts surrounding the lesion. Interestingly, we found differences between masses and calcifications in lesion conspicuity and the clarity of lesion border. For masses (n = 33), conspicuity and border clarity significantly improved using HBI method, while calcifications (n = 11) were found to be equally conspicuous with no difference in border visibility by the methods. This is mainly due to the high contrast that calcifications naturally exhibit compared to soft tissue lesions. In literature, the detection of calcifications was initially considered as one of the major limitations of DBT. Reducing the image noise level and the extent of artifacts surrounding the calcs might be advantageous here, especially for smaller and/or lower contrast lesions (Figure 5). Since there were only few patients with microcalcifications in this study, the majority of calcifications evaluated were marcocalcifications (Table 1). Further studies are needed,^6,31,32 that especially focus on the evaluation of micro- compared to macrocalcification, to proof our trend, that HBI reduces the artifacts surrounding the lesions but not impacts the border visibility.

We could not draw any conclusions about architectural distortions because the sample size was low with only one case. However, the image in that case followed the trend seen for masses, with improvement in image quality in the HBI reconstruction compared to that of FBP.

The potential reduction of artifacts using HBI reconstruction while at the same time reducing image noise level is a very important finding (Figures 2–6). In-plane artifacts typically surround a lesion and can appear as darker or brighter signals, depending on the attenuation of the surrounding tissue. These artifacts typically occur along the scan direction of the DBT system and are more pronounced for structures of higher contrasts such as large calcifications or metal clips. Artifacts surrounding the lesions may to a certain degree act as enhancers. Thus, it has been suggested they might be beneficial at the initial detection.³³ However, it is irrevocable, that these artifacts represent erroneous signals that mainly disturb the perception of the lesion’s morphology and the visibility of surrounding structures^33,34 and therefore obscure clinically relevant findings adjacent to the lesion.³⁵ Using HBI reconstruction should increase lesion detectability and improve lesion perception, which lead to higher cancer detection rates.

HBI reconstruction also significantly improved the clarity of borders and conspicuity of breast masses, which should allow easier differentiation between benign and malignant lesions. Improvement in this image parameter could lead to a reduction in the recall rate and a higher positive-predictive value for biopsy recommendation.

Invasive lobular carcinoma (ILC) is the second most common microscopic type of invasive breast carcinoma, and it is known to be difficult to detect with either standard full-field digital mammography (FFDM) or DBT.^6,7 However, a number of studies have shown increased detactability of ILC on DBT compared to FFDM, and we also found improved image quality for ILCs (Figure 4), which should help to increase its’ detection rate further.

Intrestingly, visibility of fibroglandular breast tissue was significantly improved with HBI reconstructed images in the context of a significantly lower BI-RADS breast density score (Figures 7 and 8). Lower breast density increases image sensitivity and reduces the need for additional imaging (e.g. through breast MRI). Breast density is usually graded qualitatively on a visual basis during clinical routine, and this evaluation has high inter-reader variability. In this study, the readers had a fair level of agreement about breast density for both reconstruction types. It should be noted that this study included difficult diagnostic cases with a higher frequency of dense breasts, which may have caused further inter-rating variability compared to the general screening population.

Most experts agree that the risks associated with the radiation dose of mammography are negligible in a curative clinical setting.^36,37 However, the screening situation is different in that a very large population of healthy females are exposed to radiation. Therefore, any increase in risk, even if small, has to be taken seriously.^36,37 By significantly increasing the image quality, our results indirectly suggest that the HBI method allow the use of lower radiation dosages while maintaining image quality.¹² It would be interesting to examine explicitly how much the radiatiation dose to the breast could be reduced by using the HBI method. Alternatively, it may allow more scan projections to be taken at a corresponding dose level yielding a more complete three-dimensional volume of the breast. However, acquiring more projections may be constrained by the stationary position of the detector.

A potential limitation of this study is the descriptive nature of the analysis. A side-by-side-analysis is a very direct type of evaluation but it is usually not blinded. Even if the radiologists were unaware of lesion-specific details such as histopathology and the type of reconstruction used, recognizing typical characteristics between reconstruction methods may cause a subjective preference among the readers. However, since the evaluation tasks were very specific and the readers agreed (Table 2) such an effect was probably negligable. In contrast, the breast density evaluation was performed in a blinded manner with a wash-out period of at least one and a half week in between the reading sessions and the cases were displayed in random order.²⁷ Thus, we do not expect relevant memory effects.²⁷ Finally, the study material itself included a limited number of patients, although a relatively large number of pathologies were evaluated (n = 44).

The results of lower HBI-breast density classifications have implications in the subjective risk assessment of breast cancer in denser breast, and risk assessment may need to be revised with regards to the specific reconstruction method used. Future studies could compare reconstructed HBI-assessed breast density with breast density assessed on MRI, which is often regarded as the gold-standard for measuring breast density.

In conclusion, HBI significantly improves image quality and lesion visibility compared to FBP. HBI-visibility of fibroglandular breast tissue increased while breast densities were rated lower. Applying the HBI algorithm may improve the diagnostic performance of DBT and decrease the need for additional imaging in patients with dense breasts. This may lead to higher cancer detection rates and a reduced number of unnecessary biopsies. It may result in patient dose savings, which is of special relevance in the screening context.