Lesion characterization in spectral photon-counting tomosynthesis
Bjorn Cederstrom, Erik Fredenberg, Karl Berggren, Klaus Erhard, Mats Danielsson, Matthew Wallis
This is the submitted manuscript of:
Cederström, B., Fredenberg, E., Berggren, K., Erhard, K., Danielsson, M. and Wallis, M. , “
Lesion characterization in spectral photon-counting tomosynthesis ,” Proc. SPIE
The published version of the manuscript is available at: https://doi.org/10.1117/12.2253966 See also:
Fredenberg, E., Dance, D.R., Willsher, P., Moa, E., von Tiedemann, M., Young, K.C. and Wallis, M.G., 2013. Measurement of breast-tissue x-ray attenuation by spectral mammography: first results on cyst fluid. Physics in Medicine & Biology, 58(24), p.8609. https://doi.org/10.1088/0031-9155/58/24/8609
Erhard, K., Kilburn-Toppin, F., Willsher, P., Moa, E., Fredenberg, E., Wieberneit, N., Buelow, T. and Wallis, M.G., 2016. Characterization of cystic lesions by spectral mammography: results of a clinical pilot study. Investigative radiology, 51(5), pp.340-347. https://doi.org/10.1097/RLI.0000000000000246 All publications by Erik Fredenberg: https://scholar.google.com/citations?hl=en&user=5tUe2P0AAAAJ esion characterization in spectral photon-countingtomosynthesis
Bj¨orn Cederstr¨om a , Karl Berggren a,b , Klaus Erhard c , Mats Danielsson b , Erik Fredenberg a , andMatthew Wallis da Mammography Solutions, Philips, Torshamnsg. 30A, 164 40 Kista, Sweden b Dept. of Physics, Royal Inst. of Technology (KTH), 106 91 Stockholm, Sweden c Philips Research Laboratories, R¨ontgenstrasse 24–26, 22335 Hamburg, Germany d Cambridge Breast Unit, Addenbrooke’s Hospital, Cambridge, United Kingdom
ABSTRACT
It has previously been shown that 2D spectral mammography can be used to discriminate between (likely benign)cystic and (potentially malignant) solid lesions in order to reduce unnecessary recalls in mammography. Onelimitation of the technique is, however, that the composition of overlapping tissue needs to be interpolated froma region surrounding the lesion, and this uncertainty could potentially be reduced using the 3D information fromspectral tomosynthesis. We present a first investiagtion of such an application. A phantom experiment wasdesigned to simulate a cyst and a tumor, where the tumor was overlaid with a structure that made it mimic acyst. In 2D, the two targets appeared similar in composition, whereas spectral tomosynthesis revealed the exactcompositional difference. However, the loss of discrimination signal due to spread out from the plane of interestwas of the same strength as the reduction of anatomical noise. A test on clinical tomosynthesis images of solidlesions was inconclusive and more data, as well as refinement of the calibration and algorithm, are needed.
Keywords:
Mammography, tomosynthesis, spectral imaging, photon counting, lesion characterization
1. INTRODUCTION
Solitary well defined mass lesions are a common mammographic finding, which contributes approximately 20%of overall recalls at screening. A large number of these lesions are found to be simple cysts when assessedwith ultrasound and do not require further clinical evaluation; cancer rates in solid probably benign lesions areless than 2%. Improving lesion characterization at screening and thereby lowering the number of recalls forsimple cysts would be desirable to reduce both the costs of the screening program as well as patient anxiety.Spectral x-ray imaging is an emerging technology that measures the energy dependence of x-ray attenuation. Onepotential application of spectral imaging is to characterize breast lesions identified in mammographic screeningwith the aim to reduce the number of recalls for cysts. We have previously shown in specimen experimentsthat it is feasible to discriminate between cyst fluid and solid tissue using 2D spectral mammography. Theseresults motivated a clinical pilot study for discriminating between cystic and solid lesions, which demonstrateda specificity (correctly classified cysts) of about 50% at the 99% sensitivity (correctly classified solids) level. Eventhough these results are encouraging, there is room for improvement. One limitation of lesion characterizationin 2D is that the breast composition above and below the lesion needs to be estimated by interpolation from aregion surrounding the lesion, assuming that there are no strong local changes in the composition.The aim of this investigation was to (1) demonstrate that lesion characterization can be done with spectraltomosynthesis, and (2) investigate whether the 3D information can reduce the uncertainty from the interpolationof surrounding tissue. It should be acknowledged, though, that tomosynthesis, being limited in the angular span,is not a quantitative technique like computed tomography, and that e.g. uncertainty of the local breast thicknesswill play an important role also in this case.During previous studies of 2D lesion characterization, a number of potentially limiting factors have beenidentified: Send correspondence to Bj¨orn Cederstr¨om. E-mail: [email protected]
Uncertainty in tissue composition – This factor will remain the same in tomosynthesis. Note though thatthe clinical results in Ref. indicate that some of the spread in tumor tissue composition reported in Ref. is likely due to sample preparation effects. • Statistical noise – On a global scale this will remain the same in tomosynthesis, since the dose will bethe same. However, local estimates in the 3D volume may suffer more from statistical noise than localestimates in the projection images. • Calibration imperfections – This should ideally be the same for 2D and tomosynthesis, but may differdue to different implementation of calibration procedures. This is the first 3D implementation and thealgorithm may need refinement. • Errors in interpolation of breast glandularity (structural overlap) – Here there is potential for improvementin tomosynthesis, with less structural overlap, better local estimation of glandularity, and a smaller regionin which interpolation from the surrounding is necessary. • Errors in interpolation of breast thickness – Due to the limited tomographic angle, tomosynthesis does notnecessarily have any benefit here.
2. MATERIALS AND METHODS2.1 Spectral photon-counting tomosynthesis clinical prototype
The Philips MicroDose S0 spectral tomosynthesis system (Philips Mammography Solutions, Solna, Sweden) isa not commercailly available clinical prototype. The system comprises an x-ray tube, a pre-collimator, and animage receptor, all mounted on a rigid arm (1, left). The image receptor consists of photon-counting silicon stripdetectors with corresponding slits in the pre-collimator (1, right). To acquire an image, the arm is rotated arounda point below the patient support so that the detector is scanned across the object. Each detector line viewseach point in the object from a unique angle, which results in a data set that can be used for 3D reconstruction.The width of the detector and the ppoint of rotation-to-detector distance yield a tomographic angle of about11 o . θ Tomo-graphicangle
Figure 1.
Left:
Schematic of the Philips MicroDose S0 prototype tomosynthesis system. Note the indication of thetomographic angle, which is the angle at which the outermost rays cross at a certain point in the image volume.
Right:
The photon-counting spectral image receptor and electronics.
Photons that interact in the detector are converted to pulses with amplitude proportional to the photonenergy. Virtually all pulses below a few keV are generated by noise and are therefore rejected by a low-energyhreshold. A high-energy threshold sorts the detected pulses into two bins according to energy, which enablesspectral imaging.
See Fig. 2 for a schematic of the various calibration and reconstruction steps. The data from the two energybins ( high energy bin, and sum of low and high energy bins) are first flat-field calibrated independently. Theraw photon counts are mapped to equivalent PMMA thickness using a look-up table from calibration using aPMMA step-wedge phantom. Each detector channel and energy bin is calibrated individually. This calibrationstep is performed on a regular basis to account for effects such as x-ray tube drift and detector and electronicstemperature variations. Each energy bin is then separately iteratively reconstructed to image stacks expressed inPMMA thickness. The reconstruction algorithm (SART) was modified to yield a globally linear behavior. Theimages were binned to 0.4 x 0.4 x 2 mm voxel size, in order to reduce noise and since fine-scale structures areunimportant for lesion characterization.The next step is the material decomposotion into two basis materials, aluminum (Al) and polyethylene (PE). Figure 2.
For most natural body constituents at mammographic x-ray energies, it is fair to ignore absorption edges.
6, 7
X-ray attenuation is then made up of only two interaction effects, namely photoelectric absorption and scatteringprocesses. Assuming known incident spectrum ( q ,Φ( E )) and known detector response for the two energy bins(Γ lo ( E ) and Γ hi ( E )), acquisitions over two different energy ranges yield a non-linear system of equations with aunique solution for two different material thicknesses d and d with known linear attenuation coefficients ( µ , µ ): n lo = q (cid:90) Φ dE (1)Measurements at more than two energies yield an over-determined system of equations under the assumption ofonly two independent interaction processes, and would, in principle, be redundant.As it is only possible to discriminate between two different materials in spectral imaging, the composition ofoverlapping (adipose and glandular) tissue needs to be known in order to discriminate between cystic and solidesions. In 2D, the breast thickness and glandular fraction is estimated from a region surrounding the lesion andinterpolated into the region of the lesion. Even though the depth information in tomosynthesis is limited and isreduced with increasing lesion size, the interpolated volume will still be smaller compared to 2D. A phantom experiment was carried out for the purpose of (a) assessing the 3D decomposition in terms ofhomogeneity and noise, and (b) to demonstrate that the use of 3D information can potentially improve lesioncharacterization. Two aluminum (Al) markers with identical total thickness, but one that was distributed indepth were arranged with 50 mm of tissue-equivalent material (CIRS Inc., Norfolk VA) as background tissue (Fig.3). The difference in Al thickness was chosen such that the two markers would simulated a cyst and a tumor,respectively. Note that we need further assumptions in order for this to be the case. If we assume homogeneousbreast tissue of 50% glandularity and disregard the skin, the case with 250 µ m Al represents a 9.3-mm thick cystin a 50.39 mm thick breast, whereas the case with 200 µ m Al represents an 8.7-mm thick tumor in a 49.88 mmthick breast.
50 mm CIRS 50/50250 µ m Al (tumor) 200 µ m Al (cyst)50 µ m Al Figure 3. Schematic (left) and photograph (right) of the phantom experiment. Two 250-m Al markers arranged with a50-mm slab of tissue-equivalent material as background tissue. The right marker is distributed in depth so that 200 m Alis on top of the tissue-equivalent material and 50 m is below the material.
Data from a clinical study of photon-counting spectral tomosynthesis at ImageRive, Geneva, Switzerland thatwas conducted early 2016 recently became available for this study. Symptomatic patients were examined with aMicroDose S0 system. All patients were asked to provide written informed consent prior to the examination andthe study was approved by SwissEthics. An initial analysis of the data is presented here and we plan to extendthe analysis to the final version of the study to be presented at the conference.
3. RESULTS3.1 Phantom measurements
Figure 4 shows results of the phantom experiment in 2D, i.e. as a sum over the slices. The thickness differencebetween the two targets in the Al image was 15 m. The difference in thickness was likely caused by the slightmisalignment between the 200 m Al on top of the phantom and the 50 m Al below the phantom, which is revealedby the conventional x-ray image.Figure 5 shows results in 3D as a function of slice. The left target exhibits a stronger signal at the top ofthe bottom of the phantom. The spread in the depth direction depends on target size, but as the targets inthis case are of equal size, the spread is equal and the Al thickness was normalized to the known 250 m in theleft target to generate the plot on the right-hand side with quantitative values on Al thickness. The thicknessof the left target was 200 m in the top slice, which is the expected value. The target thicknesses in the bottomslice are slightly higher than expected because the point-spread functions of the top targets extend through theentire volume and interfere with the measurement. The difference between the left and the right target was 30m, which is slightly less than the expected 50 m. Figure 6 shows results from an example case in the clinical (mm) y ( mm )
2D projection (norm. )
20 40 60 80 100 120 14020406080100 0.8 x (mm) y ( mm ) Uppermost slice
20 40 60 80 100 120 14020406080100 0.030.040.050.060.07 x (mm) y ( mm ) Lowermost slice
20 40 60 80 100 120 14020406080100 0.030.040.050.060.07
Figure 4. data set with a solid lesion (phyllodes tumor). The size of the lesion is similar in size as the Al target in thephantom experiment and the spread through the slices is similar in appearance.
Al image
50 100 15020406080100120140160 0.811.21.41.61.8 2.3 2.35 2.4 2.45 2.5 2.55 2.6x 10 sum L h i / L s u m PE image
50 100 15020406080100120140160 455055 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.84849505152535455 Al P E Al (mm) = 1.515PE (mm) = 51.52Al (mm) = 1.520PE (mm) = 51.57Al (mm) = 1.295PE (mm) = 51.17
Figure 5.
The blue line in Figure 6, right shows the expected combination of Al and PMMA for a solid lesion of anythickness derived from specimen measurements [5]. The spectral tomosynthesis volume was decomposed intoequivalent Al and PMMA thicknesses, and the lesion was evaluated in 2D (all 32 slices) and in 3D slice 27 thatruns through the lesion. The 3D evaluation yields a result that is closer to the expected attenuation. It shouldbe noted that these results are based on only one case and are so far uncertain. A more thorough investigationof a larger number of clinical cases is planned for the final study.
4. DISCUSSION5. CONCLUSIONS
Phantom experiments showed that spectral lesion characterization is improved by 3D information from tomosyn-thesis. A preliminary investigation of clinical spectral tomosynthesis data also indicated an improvement when
Al image
50 100 15020406080100120140160 0.20.250.30.35 4400 4600 4800 5000 5200 54000.8040.8060.8080.810.8120.8140.816 L sum L h i / L s u m PE image
50 100 15020406080100120140160 9.51010.5 0.2 0.25 0.3 0.35 0.49.81010.210.410.610.8 Al PE Mean Al (mm) = 0.3194Mean PE (mm) = 10.21Mean Al (mm) = 0.3223Mean PE (mm) = 10.32Mean Al (mm) = 0.2536Mean PE (mm) = 10.13
Figure 6.Figure 7. including 3D information. Improved lesion characterization holds the potential for reduction of unnecessaryrecalls, which would benefit individual women and the society as a whole.
REFERENCES [1] Erhard, K., Kilburn-Toppin, F., Willsher, P., Moa, E., Fredenberg, E., Wieberneit, N., Buelow, T., andWallis, M., “Characterization of cystic lesions by spectral mammography: Results of a clinical pilot study,”
Invest. Radiol. , 340–347 (2016).[2] Erhard, K., Fredenberg, E., Homann, H., and Roessl, E., “Spectral lesion characterization on a photon- P E d i ff . v s . b a c k g r o u n d ( mm ) Synthetic-2D
Slab 1 Slab4Lesion in focus P E d i ff . v s . b a c k g r o u n d ( mm ) Figure 8. counting mammography system,” in [
Medical Imaging 2014: Physics of Medical Imaging ], Whiting, B.,Hoeschen, C., and Kontos, D., eds.,
Proc. SPIE (2014).[3] Fredenberg, E., Dance, D., Willsher, P., Moa, E., von Tiedemann, M., Young, K., and Wallis, M., “Measure-ment of breast-tissue x-ray attenuation by spectral mammography: first results on cyst fluid,”
Phys. Med.Biol. , 8609–8620 (2013).[4] Fredenberg, E., Kilburn-Toppin, F., Willsher, P., Moa, E., Danielsson, M., Dance, D., Young, K., and Wallis,M., “Measurement of breast-tissue x-ray attenuation by spectral mammography: solid lesions,” Phys. Med.Biol. , 2595–2612 (2016).[5] Erhard, K. and van Stevendaal, U., “A simulation study on spectral lesion characterization,” in [ IWDM2016 ], et al. , A. T., ed., LNCS , 601–608 (2016).[6] Alvarez, R. and Macovski, A., “Energy-selective reconstructions in x-ray computerized tomography,”
Phys.Med. Biol. , 733–744 (1976).[7] Lehmann, L., Alvarez, R., Macovski, A., Brody, W., Pelc, N., Riederer, S., and Hall, A., “Energy-selectivereconstructions in x-ray computerized tomography,” Med. Phys.8