AAPM 2012 Annual Meeting Digital breast tomosynthesis: basic understanding of physics principles James T. Dobbins III, Ph.D., FAAPM Director, Medical Physics Graduate Program Ravin Advanced Imaging Laboratories Departments of Radiology, Biomedical Engineering, and Physics and Medical Physics Graduate Program Duke University Medical Center Acknowledgments and financial disclosure James Dobbins: Grant support from NIH (ROI CA80490) and GE Healthcare. Patent jointly held by GE and Duke; JTD is co-inventor. Has spoken at GE sponsored events. Unpaid participant at GE Medical Advisory Board mtgs. Duke University: Grant support from Siemens. FDA statement: discussion will include off-label uses and applications and devices not yet approved 1
Tomosynthesis: section imaging from multi-projection image reconstruction (limited angle tomography) Tao Wu et al, Med Phys 30(3), 2003 Breast tomosynthesis with FFDM detector Niklason et al 1997 Image from: Niklason LT et al: Radiology 205:399-406, 1997. 2
Advantages of tomosynthesis Improves conspicuity by removing overlying structures Permits section imaging with high resolution in reconstruction plane Easily performed on the high volume of mammography patients Potentially lower cost than breast CT Breast Tomosynthesis Human subject examples Subject 48 normal screen Courtesy of and copyright by Joseph Lo, PhD 3
skin, surface blood vessels, subq fat pockets Courtesy of and copyright by Joseph Lo, PhD intramammary lymph nodes Courtesy of and copyright by Joseph Lo, PhD 4
lymph nodes Courtesy of and copyright by Joseph Lo, PhD pectoralis Courtesy of and copyright by Joseph Lo, PhD 5
Subject 158 - mammo miss of cancer Courtesy of and copyright by Joseph Lo, PhD Courtesy of and copyright by Joseph Lo, PhD 6
Geometry of tomosynthesis image acquisition Geometries of motion Parallel path Partial isocentric Isocentric Chest tomosynthesis Breast tomosynthesis CBCT, RadOnc tomosynthesis 7
Breast tomosynthesis clinical implementation Parameters: up to 49 projection images up to ~ ±25 0 angle range compression paddle to keep objects still Tomosynthesis image formation 8
Tomosynthesis algorithms Shift-and-add ART (algebraic reconstruction techniques) Tuned aperture computed tomography (TACT) Iterative methods (MLEM) Matrix inversion tomosynthesis (MITS) Filtered backprojection (FBP) Shift-and-add reconstruction (simple backprojection) Acquisition geometry Shift-and-add image formation 9
The importance of deblurring Conventional tomo section After deblurring Matrix Inversion Tomosynthesis (MITS) 10
Removing the blur with MITS Direct solution using linear algebra and the known acquisition geometry Much faster computationally than iterative deblurring Better performance at narrow tube angles than filtered backprojection However. susceptible to noise at the lowest spatial frequencies ( < ~ 0.1 cycles/mm) Removing the blur with MITS Matrix form (freq space) Rewritten Solving for true structures " $ $ $ $ # T 1 T 2! T n % " F 11 F 12 " F 1n % " ' $ ' $ ' F = $ 21 ' ( $ ' $! # ' $ ' $ ' $ & # & # F n1 T = M " S F nn s = FT "1 ( M "1 # T) S 1 S 2! S n % ' ' ' ' & 11
Filtered backprojection Filtered backprojection methodology FT Spatial frequency Acquire projection images; take Fourier transform Multiply by ramp filter Multiply by roll-off filter Reconstruct by shift-and-add 12
Projection/Slice Theorem Courtesy of and copyright by James Mainprize, PhD Simple Backprojection Back Projection Reconstructed Fourier Original Fourier Very blurry. +Noise tolerant Courtesy of and copyright by James Mainprize, PhD 13
Principles of FBP for cone-beam imaging Filtered Backprojection Intuitive Interpretation Backprojection causes blur Correct the blur with an inverse filter MTF(k) Inverse filter=1/mtf(k) Ramp Filter Reconstruction Filter Courtesy of and copyright by James Mainprize, PhD Filtered Backprojection Back Projection Reconstructed Fourier Original Fourier Courtesy of and copyright by James Mainprize, PhD 14
Filtered Backprojection Ramp filter provides exact reconstruction when: Noise free Sufficient samples (no missing data) Courtesy of and copyright by James Mainprize, PhD Tomosynthesis (Limited View / Limited Angle) Back Projection Reconstructed Fourier Original Fourier Courtesy of and copyright by James Mainprize, PhD 15
Iterative reconstruction strategies Iterative Reconstruction Algorithm Z Y X - Breast volume is sampled using a three-dimensional matrix of elements (voxels) - Typical voxel size: 0.1 mm 0.1 mm 1 mm - The value of a voxel is the linear x-ray attenuation coefficient µ of that element Courtesy of and copyright by Tao Wu, PhD 16
Maximum Likelihood Expectation Maximization (ML-EM) Initial 3D Model µ (0) µ (n) Update Forward projection Measured Projections: Y Calculated Projections: Y (n) Δµ (n+1) Optimized Likelihood Function µ (end) Courtesy of and copyright by Tao Wu, PhD ML-EM Reconstruction: Likelihood Function Likelihood Function L=P(Y µ): probability of getting the measured projections Y, given a 3D model µ µ (n) is updated iteratively so that L (n+1) > L (n) The reconstruction solution is the 3D attenuation distribution model that maximizes L Courtesy of and copyright by Tao Wu, PhD 17
Optimization of acquisition parameters Key design criteria for tomosynthesis Must have effective blur removal for outof-plane structures Must handle between-plane structures reasonably Rapid recon time highly desirable Ability to use narrow tube angles desirable 18
Optimization of acquisition parameters Three parameters: Number of projection images, holding total dose constant Angle of tube movement Number of reconstructed planes Goals: Produce well-defined slice images Reduce out-of-plane blur artifacts Minimize image noise Tomosynthesis impulse response function 20-degrees of tube movement Conv Tomo MITS 11 proj images 61 proj images Courtesy of and copyright by Devon J. Godfrey, PhD 19
21 PROJECTIONS 16 o total angle 61 PROJECTIONS 16 o total angle Note: Total exposure is not equivalent in this example, so disregard SNR variation Courtesy of and copyright by Devon J. Godfrey, PhD NPS measurements of MITS reconstructions Exposure x Normalized NPS ENNPS (mm 2 mr) 500 400 300 200 100 N = 21 N = 31 N = 41 N = 51 N = 61 N = 71 (a) ENNPS (mm 2 mr) 500 400 300 200 100 N = 21 N = 31 N = 41 N = 51 N = 61 N = 71 (b) 0 0 0.2 0.4 0.6 0.8 1 1.2 cycles/mm Anterior plane 0 0 0.2 0.4 0.6 0.8 1 1.2 cycles/mm Central plane 20-degree tube motion, 19 planes, 16.7 mm plane spacing Courtesy of and copyright by Devon J. Godfrey, PhD 20
Acquisition parameters Num proj images Total tube angle Plane spacing Algorithm Breast 11-49 15-50 o 1 mm FBP, MLEM, MITS, SART Dose optimization Current dose levels: 1-2 X single-view FFDM (2-4 mgy) Dose appropriate for mass or calcs or both? Need to optimize x-ray spectra; kvp, filtration Potential role of noise reduction algorithms to reduce dose Additional dose optimization needed as radiologists gain clinical experience and better define tasks 21
Translational issues remaining Breast tomo applications: Likely use: screening, diagnostic, or both? Recall and biopsy rate: clinical trial Improvement in PPV? Diagnostic decision making and pt outcomes: larger trials needed How many views: MLO, CC or both Should a standard FFDM also be taken? Tomo performance for calcs Impact of tomo on dx workups (sufficient specificity to send directly to biopsy from tomo?) Comparison of breast CT and breast tomo The main advantage of DBT over conventional mammography is: 1. Better visibility of chest wall 2. Reduced compression required 3. Better spatial resolution 4. Reduced dose 5. Better visibility of masses in dense breasts Answer: 5 Better visibility of masses in dense breasts Ref: C. M. Hakim et al, Digital breast tomosynthesis in the diagnostic environment: a subjective side-by-side review. Am J. Roentg 195:W172-W176, 2010. 22
The main limitation of DBT when compared to breast CT is: 1. Reduced resolution in depth direction 2. Reduced resolution in x-y plane 3. Increased dose 4. Higher cost 5. Reduced visualization of chest wall Answer: 1 Reduced resolution in depth direction Ref: J. M. Boone et al, Dedicated breast CT: radiation dose and image quality evaluation. Radiology 221:657-667, 2001. Deblurring in filtered back projection is provided by: 1. An apodization filter 2. A ramp filter 3. A resampling function 4. An inversion matrix 5. An iterative process Answer: 2 A ramp filter Ref: J. Dobbins and D. Godfrey, Digital x-ray tomosynthesis: current state of the art and clinical potential. PMB 48:R65- R106, 2003. 23
Review articles: Dobbins JT III, Godfrey DJ: Digital x-ray tomosynthesis: current state of the art and clinical potential. Physics in Medicine and Biology, 48:R65-R106, October 2003. PDF: stacks.iop.org/pmb/48/r65 Dobbins JT III: Tomosynthesis imaging: at a translational crossroads. Medical Physics 36:1956-1967, 2009. 24