Transcript Document
Image Reconstruction and Inverse Treatment Planning – Sharpening the Edge – Thomas Bortfeld I. II. CT Image Reconstruction Inverse Treatment Planning Syllabus 2/13 Advances in imaging for therapy (Chen) 2/20 Treatment with protons and heavier particles; tour of proton facility (Kooy) 2/27 Treatment delivery techniques with photons, electrons; tour of photon clinic (Biggs, Folkert) 3/6 Intensity-modulated radiation therapy (IMRT, IMPT) (Seco, Trofimov) 3/13 Dose calculation (Monte Carlo + otherwise) (Paganetti, Kooy) 3/20 Treatment planning (photons, IMRT, protons) (Doppke + NN) 2 Syllabus 3/27 Inverse treatment planning and optimization (Bortfeld) 4/3 Optimization with motion and uncertainties (Trofimov, Unkelbach) 4/10 Mathematics of multi-objective optimization and robust optimization (Craft, Chan) 4/17 Dose painting (Grosu) 4/24 Image-guided radiation therapy (Sharp) 5/1 Special treatment techniques for moving targets (Engelsman) 3 Planigraphy, Tomosynthesis “Verwischungstomographie” x-ray source x-ray tube focus slice object film film Ziedses des Plantes (Netherlands), 1932 4 Computerized Tomography Greek: • tomos = section, slice • graphia = to write, to draw tomograph = slice drawer 5 Nobel prize 1979 Drs. Hounsfield and Cormack Sir Godfrey N. Hounsfield (Electrical Engineer) EMI Allan M. Cormack (Physicist) South Africa, Boston 6 First CT scanner prototype Hounsfield apparatus 7 CT “generations” 1st generation 2nd generation translation-rotation scanners 8 CT “generations” 3rd generation (fan detector) 4th generation (ring detector) Rotation-only scanners 9 Imatron: Electron beam CT (EBT) 10 Imatron: Electron beam CT (EBT) 11 Helical (“spiral”) CT Trajectory of the continuously rotating X-ray tube Start of spiral scan Table motion 12 From transmission to projection A.M. Cormack 13 Projection 14 Sinogram, “Radon” transform, (Johann Radon, Vienna, 1917) p Object f ? f Sinogram p 15 Backprojection 16 Backprojection Shepp&Logan phantom “Reconstruction” by backprojection 17 Backprojection 18 Backprojection Projections of point object from three directions Back-projection onto reconstruction plane 19 Backprojection Object Image f(r) 1/r f(r) Profile through object Profile through image Terry Peters (Robarts Institute) 20 Backprojection This is a convolution (f * h) of f(r) with the point spread function (PSF) h=1/|r| 21 What is Convolution? 22 Convolution theorem • Periodic sine-like functions are the Eigenfunctions of the convolution operation. • This means, convolution changes the amplitude of a sine wave but nothing else. • Hence, convolution is completely described by the transfer function or frequency response (Fourier transform of the PSF), which determines how much amplitude is transmitted for different frequencies 23 Convolution theorem 24 r-filtered layergram reconstruction 1. Backproject measured projections, and integrate over f 2. Fourier transform in 2D 3. Multiply with distance from the origin, |r|, in the frequency space 4. Inverse Fourier transform 25 Central slice theorem Filtered backprojection 27 Deconvolution (filtering) in the spatial domain Ramp filter in the frequency domain: sampling interval Transform back to get filter in spatial domain: 28 Deconvolution (filtering) in the spatial domain Sample at discrete points p = nDp: sampling interval Filter of Ramachandran and Lakshminarayanan (Ram-Lak) 29 Ram-Lak filter negative components 30 Shepp&Logan filter 31 Backprojecting filtered projections Terry Peters (Robarts Institute) 32 Filtered backprojection Shepp&Logan phantom “Reconstruction” Filtered backprojection by backprojection 33 Filtered Backprojection Simple backprojection Filtered backprojection 34 Reconstruction from fan projections k+l = 7: parallel k: counter of source positions l: counter of projection lines 35 Summary CT image reconstruction 1. In CT we measure projections, i.e., line integrals 2. The set of projection lines from all directions is called Radon transform or sinogram 3. Backprojection leads back into image space but introduces severe 1/r blurring 36 Summary CT image reconstruction, cont’d 4. Image can be de-blurred with deconvolution techniques 5. Deconvolution can be done in projection space using the central slice theorem 6. A common filter function is the Ram-Lak filter 7. Filters have negative components (“eraser”) to remove blurring 37 Image Reconstruction and Inverse Treatment Planning – Sharpening the Edge – Thomas Bortfeld I. II. CT Image Reconstruction Inverse Treatment Planning Brahme, Roos, Lax 1982 Source Phantom Target OAR Dose 40 The idea of IMRT "Classical" Conformation Intensity Modulation Treated Volume Treated Volume Target Volume Tumor Tumor OAR OAR Target Volume Collimator 41 “Inverse” treatment planning "Conventional" Planning Treated Volume Target Volume OAR Inverse Planning Target Volume Treated Volume OAR Collimator 42 Computer Tomography Conformal Radiotherapy Projection Intensity Modulation Radiation Source Detectors Radiation Source Target Image Reconstruction (Filtered Backprojection) Conformal Radiotherapy (Filtered Projection) Density Distribution of the Tissue Set of Prescribed 2D Dose Distributions x-ray Projection (CT-Scanner) Projection (Computer) 1D Filtering of the Projections 1D Filtering of the Projections Backprojection IMRT with Filtered Projections Set of 2D Slice Images Dose Distribution 44 G. Birkhoff: On drawings composed of uniform straight lines Journ. de Math., tome XIX, - Fasc. 3, 1940. Negative Intensities After filtering: Intensity - x 46 The inverse problem has no solution! Consider the inverse problem as an optimization problem. Define the objectives of the treatment and let the computer determine the parameters giving optimal results. Optimization basics Mathematical optimization: Minimization of objective functions actual dose prescribed dose Objective Function F(x) = Si (di - pi)2, di = f(x1,.., xn) F(x) = NTCP (1-TCP) Constraints di < dtol xi > 0 DVH constraints NTCP < 5% Parameters x = (x1, . . ., xn) (e.g., intensity values) “Decision variables” • • • • • Intensity profiles Beam weights, segment weights Beam angles (gantry angle, table angle) Number of beams Energy (especially in charged particle therapy) • Type of radiation (photons, electrons, ...) The “standard model” of inverse planning minimize N F (b ) wi d i (b ) Pi i 1 d i (b ) Dijb j dose values 2 beam intensities j all the physics is here Dij : dose contribution of pencil beam j to voxel i bj 0 50 Dose-volume histogram for OAR small weight (w) DVH Volume Volume DVH large weight (w) max D Dose max D Dose Critical structure (organ at risk) costlet ( N R ,k FR ,k (b ) wk d i ,k b Dmax, k i 1 weight importance “penalty” dose at voxel i in OAR k 2 tolerance dose x for x 0 x 0 otherwise 52 Dose-volume histogram for the target Volume DVH min D max D Dose The “standard model” of inverse planning Minimize F wTarget FTarget wRisk1 FRisk1 wRisk2 FRisk2 objectives, costlets, indicators weights, penalties, importance factors 54 The “standard model” of inverse planning • • • • High-dimensional problem: Ray intensities bj: 10,000 Dose voxels di: 500,000 Dij matrix: 10,000 x 500,000 entries 20 GByte 55 Optimization with gradient descent F(x) local min. x3 x2 x1 x0 x global min. xt 1 xt grad F ( x ) xt dF 1D: xt 1 xt dx 56 Optimization algorithms • Projecting back and forth between dose distribution and intensity maps 57 Volume effect Whole lung: 18 Gy 50% of lung: 35 Gy 78 Volume effect Power-law relationship for tolerance dose (TD): TD(1) TD(v) n v n small: small “volume effect” n large: large “volume effect” 79 Volume effect -> EUD, Power-Law (a-norm) Model Mohan et al., Med. Phys. 19(4), 933-944, 1992 Kwa et al., Radiother. Oncol. 48(1), 61-69, 1998 Niemierko, Med. Phys. 26(6), 1100, 1999 1/ a a EUD vi Di i Examples: “a-norm” (a=1/n) a 1 : EUD D a : EUD Dmax 80 Equivalent Uniform Dose (EUD) Volume [%] 100 75 Lung: EUD = 25 Gy 50 Spinal Cord: EUD = 52 Gy 25 0 0 EUD = The homogeneous dose that gives the same clinical effect 20 40 60 Dose [Gy] 80 100 81 Volume DVH constraint max V max D Dose 82 Penalties/ Weights/ … Dose, Dose-Volume Constraints 83 Summary inverse treatment planning 1. Intensity-modulated radiation therapy (IMRT) uses non-uniform beam intensities from various (5-9) beam directions 2. “Inverse planning” is the calculation of intensities that will give the desired spatial dose distribution 3. CT reconstruction techniques cannot be (directly) applied here because we cannot deliver negative intensities 4. Today “inverse planning” is usually defined as an optimization problem, which is “solved” with gradient techniques 84