OPTIMAM Image Simulation Toolbox -Recent Developments and Ongoing Studies
Premkumar Elangovan1, Andria Hadjipanteli2, Alistair Mackenzie2, David R Dance2,3, Kenneth C Young2,3, and Kevin Wells1
1Centrefor Vision, Speech and Signal processing, University of Surrey,
Guildford, GU2 7XH, UK
2NCCPM,Royal Surrey County Hospital, Guildford, Surrey, GU2 7XX, UK
3Department of Physics, University of Surrey, Guildford, GU2 7XH, UK
Abstract.Virtual clinical trials (VCTs) are increasingly being seen as a viable pre-clinical method for evaluation of imaging systems in breast cancer screening. The CR-UK funded OPTIMAM project is aimed at producing modelling tools for use in such VCTs. In the initial phase of the project, modelling tools were produced to simulate 2D-mammography and digital breast tomosynthesis (DBT) imaging systems. This paper elaborates on the new tools that have recently been developed for the current phase of the OPTIMAM project. These new additions to the framework include tools for simulating synthetic breast tissue, spiculated masses and variable-angle DBT systems. These tools are described in the paper along with the preliminary validation results. Four-alternative forced choice (4-AFC) type studies deploying these new tools are underway. The results of the ongoing 4AFC studies investigating minimum detectable contrast/size of masses/microcalcifications for different modalities and system designs are presented.
Keywords: Digital breast tomosynthesis, 2D-mammography, modelling, simulation, 4AFC, simulated masses, breast phantom
1Introduction
National breast cancer screening programs play a vital role in early detection and treatment of breast cancer in Western countries. The choice of imaging modality, or modalities, to be used in such screening programs is constantly reassessed and is still a source of debate [1]. With new modalities being introduced into the market and a wide choice of system designs available within each modality, evaluation by means of clinical trials becomes challenging as this conventional approach requires timescales that cannot keep pace with the rate of technological innovation. Recently, virtual clinical trials (VCTs) have started to be used for rapid pre-clinical evaluations of new modalities using modelling tools. This approach also provides a means to study the effects of various intrinsic design variations so as to optimize the design for the best screening performance.
The CR-UK funded OPTIMAM project is producing modelling tools for use in such VCTs [2]. The early phase of the project was predominantly focused on the design of simulation tools for 2D-mammography and narrow angle DBT systems. The framework included: (1) cancer simulation models (masses [3,4] and microcalcification [5]); (2) a ray tracing tool (image acquisition); (3) image degradation model (scatter, noise and blur) [6,7] and (4) manufacturer specific post-processing tools (image processing and reconstruction). The lesion simulation models were validated qualitatively by means of observer studies involving radiologists [3,4,5]. Models of image formation were validated quantitatively using image quality metrics [2], [6,7].
In the current phase of the OPTIMAM project, these tools have been extended to include further modelling components for use in four-alternative forced choice (4-AFC) experiments. New tools include: (1) a tool for simulating 3D breast structure [8]; (2) a model for simulating spiculated lesions[9]; (3) simulation of DBT imaging using a wide-angle geometry and (4) a user interface for 4AFC studies. In this paper, we describe in detail the new additions to the OPTIMAM framework and provide some selected results of ongoing 4AFC studies [10,11] that employ these tools.
2Materials, Methods and Results
Figure 1 provides an illustration of the OPTIMAM simulation framework. New developments have been madein the simulation oftissues and imagingas described in following sections.
Fig. 1.OPTIMAM simulation framework [2]
2.1 Synthetic Breast Model
One of the important additions to the current phase of the OPTIMAM project is asynthetic breast model [8]. This allows insertion of synthetic pathology (masses or micro calcifications) in a more realistic manner, as prior work using template multiplication [3,4,5] led to realism issues [12] with intersecting anatomical features at certain anatomical locations. These issues have been particularly apparent when inserting masses into DBT image data. Using synthetic tissue allows pathology to be inserted by voxel replacement which produces greater levels of realism in the resulting images. The elements of normal breast tissueweresimulated by first extracting various anatomical features from DBT images of patients.
Fig. 2.(a)Cross-sectional slice fromthe breast model; (b) processed 2D image and (c) reconstructed DBT plane
Fig. 3.Normalized power spectrum curves for 2D (blue, red dotted lines represent 5th and 95th percentiles) [8]. Vertical dashed lines indicate window of spatial frequencies recognized as being mainly attributable to image texture rather than other factors such as quantum noise
The process starts by extracting glandular segmentsfrom DBT images using a region-growing method and subsequently decluttering the segmentsby applying a series of morphological operations. This process is repeated until a database of glandular segmentsis produced. An empty breast volume is created and then filled with fatty tissue. Glandular segmentsare then gradually added to the empty volume until the required glandularity has been achieved. A wireframe of blood vessels and Cooper’s ligaments was produced by spline interpolation of in-focus landmarks produced by manual annotation of DBT image planes. The wireframe representation was then dilated toanatomically appropriate diameters and added to the breast volume containing the glandular matrix to simulate these anatomical components with high spatial frequencies. Figure 2 shows an example of a slice through the breast model and simulated 2D and DBT images of the model.
The breast model was validated by computing the power spectraof simulated images and comparing thesewith the spectra computed for real images. The methods described by Hill et al [13] and Cockmartin et al[14] were used for calculating the power spectrum of real and simulated images. A selection of 300 3cm x 3cm regions of interest in simulated images wasused to calculate the power spectrum. The real image dataset contained 40 sets of patient images acquired using the Hologic Selenia Dimensions 3D breast tomosynthesis system (Hologic, Bedford, Massachusetts, USA). Figure 3 shows the mean power spectra of real and simulated 2D images. The curve for simulated images closely overlaps with that of real images.
Simulated 2D and DBT images of the breast model were interleaved with real images and presented to a team of experienced observers in an ROC study. The team comprised of 4 observers with an average breast screening experience of around 10 years.The average area under the curve (AUC) for 2D and DBT images were 0.53±.04 and 0.55±.07 respectively [8]. The results indicate that the observers had difficulty in differentiating real and simulated images.
2.2 Mass Simulation Model
In the early phase of the OPTIMAM project, irregular lesions were simulated by a fractal-growth process known as diffusion limited aggregation (DLA) [3,4]. The appearance of such irregular simulated masses after insertion into clinical radiological images was found to be largely indistinguishable from real masses with similar appearance using an ROC paradigm. By changing the parametric prescriptions of the DLA growth model it was possible to simulate masses of different densities and appearances. In the current phase of the OPTIMAM project, this model has now been extended to simulate spiculated lesions [9]. Such lesions are characterized by linear structures branching from a central core. In order to simulate spicules with realistic appearance, in terms of curvature and distribution, various features were extracted from patient DBT images and used as a guideline for generating synthetic spicules. This was performed by (Fig 4a) manually scrolling through DBT planes and marking the points at which spicule segments appear in-focus across a variety of different planes. 3D spline interpolation (Fig 4b) was then used to connect the selected points to produce a 3D skeleton for each spicule. The resulting 3D “spicule skeletons” (Fig 4c) were converted into thin horns (Fig 4d) and placed on the surface of a DLA mass (Fig 4e-f). Finally, the 3D geometric horns were filled up with fine fibrotic structures in order to provide the correct density and internal texture. This fibrotic characteristic can be observed in specimen X-ray images of real spiculated lesions.Some examples of simulated irregular and spiculated masses are shown in Figure 5.
Fig. 4.Spiculated mass generation process [9]
For preliminary validation, a total of 13 simulated spiculated masses were inserted into 2D and DBT patient images and presented to an experienced breast radiologist (>15 yearsscreening experience). The simulated images were interleaved with images containing real spiculated masses and presented blind to the radiologist for feedback. The radiologist rated 60% of the simulated lesions in 2D and 50% of the simulated lesions in DBT to be realistic. The radiologist was able to correctly identify all the 2D and DBT real images in the dataset.
Fig. 5.Examples of (a) simulatedirregular and (b) simulated spiculated masses
2.3 Image Degradation Model
In the early phase of the OPTIMAM project, we modelled the system acquisition and image degradation processes of narrow angle DBT systems, in particular the geometry and settings of the Hologic Selenia Dimensions 3D breast tomosynthesis system. The entire simulation framework was validated by simulating geometrically defined objects and quantitatively comparing the simulated images with real images of the same objects [2]. The process starts with the simulation of X-ray projections through a 3D voxelized phantom for a particular X-ray spectrum. Then, scatter estimated from a lookup table produced from Monte Carlo simulations [6] for different breast thickness/glandularity is added to the projections. Finally the appropriate noise and blur are added to the images using methods described in our previously published work [7].
In the present phase of the project we have extended these tools to simulate wide angle geometry such as that used by the Siemens Mammomat Inspiration (Siemens Healthcare, Erlangen, Germany) systems. The new system was characterized using standard techniques for measuring the pre-sampled MTF (modulation transfer function), NNPS (normalized noise power spectrum), STP (signal transfer properties), flat fielding correction map, focal spot size and focus motion. The increase in motion blur due to tube motion in the wide-angle DBT system was included in the simulations by elongating the focal spot along the direction of the tube motion. A detailed description of the image simulation procedure is given in our previously published work [2].
2.4 Ongoing 4AFC Studies
VCTs conducted during the early phase of the OPTIMAM project predominantly involved modification of real images to simulate the synthetic objects representative of cancers. Planned virtual clinical trials for the current phase of the project are aimed at using the 4AFC paradigm. AFC experiments can be used to examine the fundamental human signal detection response for a variety of stimuli. This paradigm has been previously used to compare and evaluate breast imaging modalities. The figure of merit for our 4AFC experiment is the minimum detectable contrast or object size which is defined as the value at which the observer makes 92.5% correct decisions (d' of 2.5) for a given experimental condition. Two studies are currently underway, details and results are given below.
The objective of the first study is to consider the minimum detectable contrast in 2D-mammography and DBT systems for simulated masses and solid spheres. Masses and spheres of 5mm diameter were simulated and inserted into synthetic breast models of different thicknesses at random depths. 2D and DBT (15º tube sweep/15 projections) images were produced for image acquisition and detector settings for the Hologic Selenia Dimensions 3D breast tomosynthesis system. Images were simulated with the inserted lesions having different contrast levels (1% to 6% in steps of .5%), resulting in 45 mass images and 15 sphere images per contrast level for both 2D and DBT. The contrast was varied by changing the attenuation properties of the inserted masses and spheres. The contrast was defined as the relative difference between background and target pixel intensity measured in the 2D image. Hologic image processing (Hologic LORAD FFDM Selenia V5.0) and reconstruction tools were used to post-process the images, which were then cropped to 3x3cm2 with the target in the center. The results of the mass/sphere 4AFC study are shown in Figure 6a[10]. They indicate that the observers needed approximately twice the signal contrast to correctly identify a mass in 2D-mammography (0.094±.005) compared with DBT (0.040±.001). The differences were much bigger when spheres were used instead of simulated irregular masses. However, the minimum detectable contrast for spheres was much lower in both 2D-mammography (0.075±.008) and DBT (0.0074±.002) compared to simulated masses. This might be because of the regular geometry of spheres making them easy to spot in DBT images.The errors given are the standard errors in the average minimum detectable contrastfrom all the observers.
The objective of the second study is to determine the minimum detectable micro calcification diameter in breast images using 2D-mammography and narrow (15º tube sweep/15 projections) and wide (50º tube sweep/25 projections) angle DBT. The microcalcification clusters (5 calcs/2.5mm cluster size) were produced containing microcalcifications of different sizes (100 μm to 500 μm in steps of 25 μm). These clusters were inserted in the middle of 6cm thick synthetic breast models and 50 images per experimental condition were produced for each modality considered. The mean glandular dose to the breast was set to approximately 2.5mGy in all three cases. Briona (Real Time Tomography, LLC, Philadelphia, USA) image processing and reconstruction tools were used to post process the images, and the images were cropped to 2.7x2.7cm2 with the target in the center. The results of the microcalcification cluster study are shown in Figure 6b [11]. The results indicate that the minimum detectable microcalcification diameters for 2D-mammography, narrow angle DBT and wide angle DBT are 164±5μm, 210±5μm and 255±4μm respectively. The errors given are the standard errors in the average minimum detectable calcification diameterfrom all the observers. Currently, the mass/sphere 4AFC study is underway for different mass sizes, and the microcalcification cluster size 4AFC study is underway for different heights and breast thicknesses.
Fig. 6.Results of the 4AFC observer study for (a) lesion and sphere targets [10] and (b) microcalcification clusters [11]
3Conclusions
Recent improvements to the OPTIMAM image simulation toolbox have been demonstrated. Much of the recent progress has come from improving the tissue/image simulation tools and adoption of the 4AFC paradigm for VCTs. The preliminary validation results of the tissue simulation model are very promising and encourage further development. The results of the 4AFC studies demonstrate the variations in detection performance for different types of targets and system designs. This highlights the potential of virtual clinical trials encouraging further studies.
Acknowledgements.This work is part of the OPTIMAM2 project funded by Cancer Research UK (grant, number: C30682/A17321).We are grateful for Hologic’s assistance with the reconstruction. The authors thank colleagues at NCCPM, Dr. Vicky Cooke at the Jarvis Breast Screening Centre, Guildford and observers at St Georges Hospital, London for invaluable assistance.
References
1. Hevie, M.A., Chang, J.T., Hendrick, R.E., Banerjee, M.: Reduction in Late-stage Breast Cancer Incidence in the Mammography Era: Implications for Overdiagnosis of Invasive Cancer. Cancer 120, 2649-2656 (2014).
2. Elangovan, P., Mackenzie, A., Diaz, O., Rashidnasab, A., Dance, D.R., Young, K.C., Warren, L.M., Shaheen, E., Bosmans, H., Bakic, P.R., Wells, K.: Development and Validation of a Modelling Framework for Simulating 2D-mammography and Breast Tomosynthesis. Phys. Med. Biol. 59, 4275-93 (2014).
3.Rashidnasab, A., Elangovan, P., Diaz, O., Mackenzie, A., Young, K.C., Dance, D.R., Wells, K.: Simulation of 3D DLA Masses in Digital Breast Tomosynthesis.Proc.SPIE 8668, 86680Y (2013).
4. Rashidnasab, A., Elangovan, P., Yip, M., Young, K.C., Dance, D.R., Wells, K.:Simulation and Assessment of Realistic Breast Lesions using Fractal Growth Models.Phys.Med.Biol.15, 5613-5626 (2013).
5. Shaheen, E., De Keyzer, F., Bosmans, H., Dance, D.R., Young, K.C., Van Ongeval, C.: The Simulation of 3D Mass Models in 2D Digital Mammography and Breast Tomosynthesis.Med.Phys. 41, 081913-1-17 (2014).
6. Diaz, O., Dance, D.R., Young, K.C., Elangovan, P., Bakic, P. R.,Wells, K.:Estimation of Scattered Radiation in Digital Breast Tomosynthesis.Phys.Med.Biol. 59, 4375-90 (2014).
7. Mackenzie, A., Dance, D.R., Diaz, O., Young, K.C.:Image Simulation and a Model of Noise Power Spectra Across a Range of Mammographic Beam Qualities. Med.Phys. 41, 12901-14 (2014).
8. Elangovan, P., Dance, D.R., Young, K.C., Wells, K.: Generation of 3D Synthetic Breast Tissue. Proc.SPIE 9783,9783081-6(2016).
9. Elangovan, P., Alrehily, F., Pinto, R.F., Rashidnasab, A., Dance, D.R., Young, K.C., Wells, K.: Simulation of Spiculated Breast Lesions. Proc.SPIE 9783, 97832E1-5(2016).
10. Elangovan, P., Rashidnasab, A., Mackenzie, A., Dance, D.R., Young, K.C., Bosmans, H., Segars, W.P.,Wells, K.: Performance Comparison of Breast Imaging Modalities using a 4AFC Human Observer Study. Proc.SPIE 9412, 94121T1-7 (2015).
11. Hadjipanteli, A., Elangovan, P., Looney, P.T., Mackenzie, A., Wells, K., Dance, D.R., Young, K.C.: Detection of Microcalcification Clusters by 2D-mammography and Narrow and Wide Angle Digital Breast Tomosynthesis. Proc.SPIE 9783,9783061-8 (2016).
12. Rashidnasab, A., Elangovan, P., Mackenzie, A., Dance, D.R.,Young, K.C., Bosmans, H., Wells, K.:Virtual Clinical Trials using Inserted Pathology in Clinical Images: Investigation and Assumptions for Local Glandularity and Noise.Proc.SPIE 9412, 94122D1-7 (2015).
13. Hill, M.L., Mainprize, J.G., Carton, A.K., Saab-Puong, S., Iordache, R., Muller, S., Jong, R.A., Dromain, C.,Yaffe, M.J.: Anatomical Noise in Contrast-enhanced Digital Mammography. Part I. Single-energy Imaging. Med. Phys. 40, 051910 (2013).
14. Cockmartin, L., Bosmans, H., Marshall, N.W.: Comparative Power Law Analysis of Structured Breast Phantom and Patient Images in Digital Mammography and Breast Tomosynthesis. Med. Phys. 40, 081920 (2013).