Development of Breast Models for Use in Simulation of Breast Tomosynthesis and CT Breast Imaging Stephen J. Glick* J. Michael O Connor**, Clay Didier**, Mini Das*, * University of Massachusetts Medical School ** University of Massachusetts, Lowell Background CT Breast Imaging (CTBI) and Breast Tomosynthesis (BT) imaging systems are currently being developed and studied by a number of researchers and commercial vendors. Preliminary evidence suggests that these tomographic breast imaging systems have potential for improving visualization of breast masses at approximately equivalent dose to mammography. Some BT and CTBI system design and acquisition parameters CsI thickness Pixel size Electronic noise levels kvp setting X-ray tube filter Reconstructed voxel size Reconstruction filter, # of iterations Magnification (focal spot blurring) Objective assessment of image quality Barrett and Myers (Foundations of Image Science), any meaningful approach to optimizing an imaging system must include definitions of 1. the specific task to be performed, 2. the observer, 3. an object model representing the objects to be imaged, and 4. a figure-of-merit used to evaluate task performance 1
3D Digital Breast Models 3D power-law noise, (Burgess et al, 2001, Reiser et al 2008) Clustered lumpy background (Rowland et al, 1992). Bakic 3D breast phantom (Bakic et al, 2002, Zhang et al, 2008) - Geometrical shapes Bliznakova 3D breast phantom (Bliznakova et al, 2003, 2010) - Geometrical shapes, power law noise etc. Li et al, 3D breast phantom (Li et al, 2009) - Based on clinical breast CT images Proposed UMASS model based on surgical mastectomy Power-law specimens noiseclustered Lumpy Background UMMS prototype bench-top CTBI system Varian Rad94 X-ray tube Bow-tie filter Rotary Stage Varian 2520 Flat- Panel Detector UMMS Benchtop CT Imaging System Pendant Breast Holder Mastectomy Imaging Varian Rad94 X-ray tube Varian 2520 Flat- Panel Detector Uncompressed Breast Holder Compressed Breast Holder Bow-tie filter Rotary Stage 2
#19 - Invasive Ductal Carcinoma Goal of Study Use CT images of mastectomy specimens to develop an ensemble of 3D breast phantoms that can be used as input to computer simulation software to generate realistic simulated CT slices Simulation Methodology Reconstruction 3D Breast Model Phantom Generation Processing Steps Simulation Reconstruction Generate Model CT Reconstruction CTBI or BT Simulation 1. Projection Averaging 2. Cupping artifact compensation 3. Anisotropic diffusion filtering 4. Segmentation - classification of voxels a. Binary method b. Fuzzy mixture method Projections 3
Projection Averaging Acquire 10 projections per angle Cupping Correction by Altunbas Method* RATP (Radial Adipose Tissue Profile) Sample Radials to determine RATP * Med. Phys., 2007. 34(7) Weighted Fit to RATP Clinical Dose = 2.3 mgy High Dose = 23.0 mgy RATP fit treated as additive noise. Assuming circular symmetry, the correction is applied to entire slice in order to mitigate intraslice variation Preprocessing Step Anisotropic Diffusion Filtering* I (p( (G I) ) I) t Example Histograms Clinical Dose = 2.3 mgy High Dose = 23.0 mgy Reconstruction After ADF Filtering *Perona and Malik et al, 1990 High Dose with ADF Filter 4
Method 1 : Classify each Voxel as Adipose or Fibroglandular Tissue (Binary Method) Histogram after ADF Filtering Method 1 : Classify each Voxel as Adipose or Fibroglandular Tissue (Binary Method) Adipose Tissue Fibroglandular Tissue reconstruction after ADF filtering Breast object model using Binary Method Method 2 : Voxels can be Weighted Mixture of Adipose and Fibroglandular Tissue (Fuzzy Mixture Method) Method 2 : Voxels can be Weighted Mixture of Adipose and Fibroglandular Tissue (Fuzzy Mixture Method) Adipose Tissue Histogram after TV Filtering Weighted Mixture Fibroglandular Tissue reconstruction after TV filtering Breast object model using Fuzzy Mixture Method 5
Using Breast Object Models to Simulate Cone-beam Projections UMMS Cone Beam Simulation Software (CBSS) - (Vedula et. al. SPIE 5030, 2003, Gong et al, Med Phys, 2006) TASMIP (tungsten anode spectral model, Boone et al, Med. Phys. 24, 1997) or other spectral models. Spectra scaled to provide specified mean glandular dose [DgN] (Thacker et. al. Phys. Med. Biol. 49, 2004) to pendant breast X-ray transport through breast modeled using Siddon s ray-tracing, and attenuation coefficients defined by Johns and Yaffe PMB 1987 Scatter component from Monte Carlo simulation Signal and noise propagation through CsI indirect detector using parameters from a serial cascade model Reconstruct projections using either Feldkamp FBP or a penalized maximum likelihood iterative reconstruction algorithm Reconstruction Validation Reconstruction -Binary Reconstruction - Fuzzy Reconstruction: Feldkamp FBP, no roll-off of ramp filter, voxel.254mm 3 Simulation parameters approximate measured technique: 40kVp, 0.5mAs, 300 projections, no external filter, approximate Mean Glandular Dose of 2.3mGy Validation - Quantitative Validation - Quantitative Reconstruction Reconstruction -Binary Reconstruction - Fuzzy Reconstruction Reconstruction -Binary Reconstruction - Fuzzy FG 0.0333 7.0% Adipose 0.0267 7.0% FG 0.0317 7.1% Adipose 0.0251 6.4% FG 0.0315 8.4% Adipose 0.0254 7.2% FG 0.0324 8.1% Adipose 0.0253 10.9% FG 0.0313 7.7% Adipose 0.0250 10.5% FG 0.0318 8.4% Adipose 0.0257 9.5% Contrast 24.8% Contrast 26.2% Contrast 24.2% Contrast 28.4% Contrast 25.0% Contrast 23.5% 6
Sagital Coronal Breast CT Simulation SPECIMEN PHANTOM SIMULATION Research Power Law Analysis Validation of β for CT slices # of Specs β σ R 2 Range Freq end Freq end Avg. Burgess 213 2.83 0.35 0.99 1.0 1.0 β CT = β Mammo - 1 P (f ) Metheany 44 1.86 0.32 0.9 0.9-2.6 0.22-0.5 0.45 O Connor 20 1.94 0.42 0.98 0.88-2.59 0.22-0.5 0.35 20 2.21 0.25 0.95 1.6-2.72 0.5 0.5 a f Freq start =.1 Generation of Compressed Breast Phantom (for DBT simulations) Reconstruction of compressed mastectomy specimen Another approach for generating compressed phantoms - use template Compressed breast phantom (binary method) Template based on CIRS Stereotactic Needle Biopsy Training Phantom 7
Mammograms from Compressed Breast Phantoms Example Usage of Breast Phantoms 98 tomosynthesis reconstructions, half with calc cluster present Two reconstruction methods compared Four observers (physicist) selected location and confidence of presence Average LROC Area Comparisons PML FBP 1 mgy aquisition 1.2 1 0.8 0.6 PML FBP 0.4 0.2 0 1.5 mgy 1.0 mgy 0.7 mgy p= 0.0007 p=0.014 p=0.028 8
Summary Two methods for generating breast objects based on high-dose CT imaging of mastectomy specimens An ensemble of uncompressed and compressed breast phantoms and simulation software can be used to explore effect of acquisition and design parameters on image quality Acknowledgements This work is supported by NIH/NIBIB EB02133 ( Feasibility of CT Mammography Using Flat-Panel Detectors ) and NIH/NCI CA102758 ("Iterative Reconstruction for Breast Tomosynthesis ). The contents are solely the responsibility of the authors and do not necessarily represent the official views of the National Institutes of Health. Patients (who consent during difficult period of their life). Clinical Team of Comprehensive Breast Clinic at Levine Cancer Center, UMass Memorial Health Care (UMMHC). Dr. Robert Quinlan, Medical Director. Pathology Department and Cytology Lab, UMass Medical School (UMMS). Dr. Ashraf Khan. UMMS Tomographic Breast Imaging Lab (TBIL) Team 3D Digital Breast Models 3D power-law noise, (Burgess et al, 2001, Metheany et al, 2007) Clustered lumpy background (Rowland et al, 1992). Bakic breast phantom (Bakic et al, 2002, Zhang et al, 2008) mammogram Proposed UMASS model based on CTBI specimens Power-law Bakic noise Clustered breast phantom Lumpy Background 9