An Investigation of Slot-scanning for Mammography and Breast CT

Abstract

Mammography and breast CT are important tools for breast cancer screening and diagnosis. Current implementations are limited by scattered radiation and/or spatial resolution. In this work, we propose and develop a slot scan-based system to be used in both mammography and CT mode that can limit scatter and collect sparse CT data for improved image quality at low radiation exposures. Monte Carlo simulations of an anthropomorphic breast phantom show a factor of 10 reduction in scattering amplitude with our slot scan-based system compared to that of a full-field detector mammography system (area mode). Similarly, slot-scan improved the MTF (particularly the low-frequency response) compared to an area detector. Investigation of sparse CT sampling with doubly sparse acquisition data return better quality reconstruction, for which our slot-scanning system is capable, over angle-only projection. Thus, a system with the combined ability for slot-scanning mammography and slot-scanning breast CT has the potential to deliver improved dose-efficient imaging performance and become viable breast cancer screening and diagnostic tools.

1. INTRODUCTION

Mammography has been the preferred screening modality for breast cancer because it is cost effective, safe and time convenient. Yet, breast cancer remains a leading cause of cancer related death among women in the United States. Moreover, mammography has led to overdiagnosis at a rate of 45%¹, causing significant economic cost in return for minimal improvement in death rates.² Mammography is limited, in part, by scattered radiation that reduces feature contrast, but also due to the overlap of anatomical features in projection imaging. Breast CT addresses anatomical overlap by creating 3D image volumes; however, it too is subject to reduced conspicuity of lesions and microcalcifications due to scatter.

Anti-scatter grids has been employed in many mammography and breast CT systems but they cause loss of primary x-rays³ and introduces artefacts that are not easily mitigated with simple gain and offset corrections due to the non-linearity of x-ray absorption.⁴ Imaging of pinhole sheets to subtract the scattered x-ray signal has also been employed but requires additional x-ray exposures.⁵ Post-processing scatter correction has been performed with Monte Carlo simulations but is computationally intensive.⁶

Slot scanning mammography was developed a decade ago⁷ and provides one way to reduce scattered radiation in measurements. Prior implementations were designed with a moving x-ray source, stationary slit, and moving detector. Such designs were found to have significant losses in spatial resolution due to mechanical instability of source motion (effectively extending the x-ray focal spot size and increasing source blur). We propose a new system with a stationary x-ray source, moving slit, and moving line detector to address this instability and to push spatial resolution limits of a slot-scanning mammography system.

Additionally, we propose a form factor that permits use of the same system for breast CT. This has the potential to bring both the scatter advantages of a slot-scanning system as well as potential resolution improvements from use of a line detector to breast CT. One difficulty with such a system is the increased time required for a data acquisition. However, such a system also presents a unique opportunity to collect novel sparse acquisition data. Specifically, most previous investigation into sparse CT acquisition have concentrated on angle sparsity (i.e., collecting whole projections at fewer rotation angles). With a line detector, there is the possibility for two-dimensional sparsity (e.g., collecting partial projections at subsets of rotation angles). Not only do such acquisitions permit alternate avenues for dose reduction (as in related SparseCT work⁸), but it also offers the potential to significantly reduce acquisition time.

This paper presents a series of investigations illustrating methods and initial performance characterizations of this proposed system. In particular, we compare scatter contributions of slot vs. area mode acquisitions in Monte Carlo studies and physical MTF measurements. We also investigate the use of the system for sparse data breast CT in a set of simulation studies where doubly sparse datasets are collected. We demonstrate this potential with comparisons of reconstructions of a breast phantom using subsets of projection angles only and both projection angles and radial bins.

2. METHODS

The slot-scanning breast imaging investigations were conducted based on a prototype test bench design illustrated in Figure 1. Both a CAD model of test bench and the physical bench are shown. It comprises of a Time Delay and Integration (TDI) detector (Shad-o-scan, Teledyne technologies, Inc.) mounted on a motorized linear stage (406XL, Parker Hannifin Corporation) that is connected via a shaft assembly (Thomson industrials, Inc.) to a custom x-ray beam slit (Ralco Medical Components, Inc.). The current implementation of the bench uses a 30 kV, 50 mA x-ray beam with a 0.52 mm Al filter (Generator: SHF-1030, Sedecal, Inc., X-ray source: M1581, Varex©) which is collimated by a slit assembly, which consist of lead shutters, to the active area of the detector. A quadrature encoder is read out from the motorized linear stage to synchronize with the TDI line rate and exposure rate of the x-ray source. The system geometry has a 650 mm source-to-detector distance and 325 mm source-to-axis distance for CT. For mammography, the breast is expected to be approximately 100 mm from the detector.

Figure 1:

(Left) CAD model of a prototype test bench system to investigate slot scanning mammography and breast CT. Shown without rotary table to illustrate the coupling between slit and detector motion. (Right) Photograph of the test bench with rotary table.

To predict the amount of scatter in the slot scanning system, we performed Monte Carlo simulations on a breast phantom generated from VICTRE (https://github.com/DIDSR/breastPhantom) with two inclusions embedded within: a 5.6 mm diameter spherical nodule and a 1.2 mm³ calcification (Fig. 2). The attenuation coefficients for each material are computed from XCOM NIST.⁹ We compare slot-scanned vs. area mode detection (a standard mammography geometry comprised of an x-ray source, object and flat-panel detector) and summarize scatter-to-primary ratios in the two cases. Monte Carlo simulations used the Monte Carlo code described in Sisniega et al.¹⁰ with 1 × 10⁵ photons/pixel.

Figure 2:

3D anthropomorphic breast phantom employed for simulating scatter profiles and performing sparse reconstructions. (a) With skin. (b) Without skin. (c) Coronal slice showing added inclusions.

To additionally quantify improvements in the physical slot-scanning system, a system MTF was measured. Again, slot- and area-modes are compared (area mode is emulated by eliminating the slit assembly). The MTF is measured from an edge response of a 50×50×5 mm³ tungsten block sandwiched between two 2.5 cm PMMA blocks – emulating scatter from compressed breast tissue.

To investigate the potential of sparse CT acquisition in the proposed system, simulation studies were performed. Specifically, rather than acquiring a standard “full-area” projection where the line detector is translated radially with a constant x-ray beam (collimated to the line by the slit assembly), we pulsed the x-ray tube to obtain projection data that is radially sparse. As a comparison, we investigate classical angle-sparsity by reducing the number of projection angles. This sparse data is reconstructed using a penalized-likelihood reconstruction using Huber function with shape parameter (δ) set to 10⁻¹⁰ (approximating a total variation penalty). We emulated a pulsed x-ray acquisition wherein our system for performing low dose CT breast imaging, we systemically performed compressed-sensing based reconstructions with varying sampling patterns of projection space. The objective function minimum is iteratively solved using a separable paraboloidal surrogates (SPS) approach.¹¹

3. RESULTS

Figure 3 illustrates MTF calculations on the test bench prototype. Horizontal (parallel to the scan direction) and vertical MTFs were computed using the edge spread from a tungsten beam-blocker (sandwiched between PMMA). We see that the MTF is significantly better in the slot-scanning mode over an area detector. For both modes, the MTF along the horizontal direction is worse than in the vertical direction and is attributed to errors in synchronization of the motorized linear stage and TDI line rate.

Figure 3:

Horizontal and vertical MTF measurements in area- and slot-scanning modes from test bench data. Note the significant low-frequency drop in MTF performance due to the increased scatter in the area-mode scans.

Figure 3 also shows a significant drop in modulation at low frequencies going from area to slot mode. This is most likely attributed to reduced scatter as suggested by Monte Carlo simulation of the breast phantom in Figure 4. The scatter distribution of the slot-scanned phantom has a similar but slightly narrower distribution than the area-scanned results. More importantly the overall magnitude of the scatter is markedly reduced with the slot scanning method – by approximately a factor of 10. Looking at the scatter-to-primary ratio in each case illustrates a variable percentage of scatter with, in general, a higher ratio in the center of the breast. Looking at the numeric values, we again find an order of magnitude reduction in scatter-to-primary: approximately 30% for the area scan and 3% for the slot-scanned system.

Figure 4:

Illustration of scatter and scatter-to-primary ratio in projections of a breast phantom in (a) area-mode acquisitions and (b) slot-scanning mammography. Note that scatter-to-primary ratio is decreased by roughly an order-of-magnitude from ~30% to ~3%.

To test the potential of our slot-scanning system for compressed sensing-based low-dose CT breast imaging, we considered several breast phantom reconstructions. Figure 5 compares three modes of acquisition: 1) filtered-backprojection with 360 “full” area projections over 360° (we denote this as 100% of projection data); 2) total-variation penalized-likelihood (PL) using subsets of “full” projection angles (e.g., 10% of data using 36 angles, and 5% using 18 angles); and (3) total-variation PL method using both angular undersampling (e.g., 72 and 36 projections) and radial subsampling where only 50% of the radial bins in each projection is sampled. This sampling used regular intervals of 11.2 mm – emulating an x-ray tube that was pulsed on and off during the translation of the slit and line detector. The combined angular and radial subsampling represents 10% or 5% of “all” projection. Thus, we may compare the performance of singly sparse data using angular-only undersampling versus doubly sparse data that is undersampled in both angle and radial bin.

Figure 5:

A comparison of reconstructions using (a) “full” data sampling; and (b)10% of the data via angular subsampling only versus (c) a combination of angular and radial subsampling. Similar comparisons using 5% of the data using (d) angular undersampling only and (e) angular and radial undersampling are also shown. While there is a consequent reduction in image quality when only 10% of projection data is acquired, the double-sparsity of the slot-scanned data shows improved contrast recovery of the small microcalcification simulated in the central upper half of the phantom. These effects are more pronounced in the cases where only 5% of the data is used. In these cases, the contrast of the larger emulated lesion in the bottom half of the phantom is also improved with doubly sparse projection data.

We note that there is a degradation in image quality from the 100% data case; however, the image quality is also similar to what one would expect for a reduced fluence acquisition. In comparing the two 10% reconstructions, we see that the doubly sparse projection data yields a slightly higher image contrast and less artefacts. In particular, note the decreased contrast of the small calcification in the central upper half of the scan. This improvement is even more pronounced when only 5% of the projection data is used.

To provide a more quantitative assessment of the improvements possible with the doubly-sparse slot-scanned data and to investigate the influence of x-ray pulse width (or equivalently the distance traveled by the line detector when the x-ray is pulsed on), we computed contrast recovery coefficients (CRCs) for the calcification and the larger spherical lesion in a series of reconstructions. This experiment is summarized in Figure 6. We note that for the 10% data scenario, doubly sparse data achieves a CRC of about 1 for the lesion over the full range of radial subsampling widths (slight above unity in some cases showing some “boosting” of contrast); whereas the singly sparse data achieves a 95% recovery. For the 5% data case, the performance gap is greater with about 90% and 72% recovery in the doubly- and singly-sparse cases, respectively. Similar improvements are seen in the case of the microcalcification; however, CRC levels are lower due to the small feature size. Moreover, there are trends in performance for the doubly-sparse data as a function of radial sampling width. Interestingly, trends are different for the two targets with the CRC maximized for the spherical lesion at the center of the range, whereas the calcification case finds better performance for the largest band widths.

Figure 6:

A comparison of contrast recovery for singly- and doubly-sparse data achievable with area- and slot-mode scanning respectively. In the slot-scanning cases, it is possible to select different x-ray pulse intervals to effectively sampled different periodic band widths in each projection. We see that the double-sparse data outperforms the singly sparse data for the two levels of data sparsity (10% and 5%). Moreover, contrast recovery can be affected by the width of the bands in the doubly-sparse slot data. Interestingly, the optimal slot width depends on the target with the microcalcification CRC higher with larger widths and the spherical lesion obtaining an optimum in the middle of the investigated range.

4. CONCLUSION

We investigated and begun the development of a novel slot-scanning system for both projection mammography and breast CT. Slot-scanning is an effective way to eliminate scattered x-rays from projection data as illustrated in simulation studies and physical measurements. Not only does this bring potential image quality improvements for projection mammography, this also has the potential to improve contrast and spatial resolution in breast CT. The proposed system permits novel doubly-sparse data acquisition which is more effective in maintaining image quality for reduced data collections. Such sparse acquisitions are important to maintain low dose protocols and provide a potential avenue for increased speed in the data collections. For example, even though the current doubly-sparse acquisition was discussed in terms of a pulsed x-ray acquisition for each projection image, there are alternate data acquisition modes where the latency of the tube in the “x-ray off” cycle is used to collect data in the next projection angle – thus, in effect, there is no “off” cycle and doubly-sparse data may still be collected. The potential speed-up depends on percentage of data required in each projection image. Ongoing work will explore variable projection percentages as well as these efficient data acquisitions.