Evolutionary optimization in (deformable) registration of medical
Transcription
Evolutionary optimization in (deformable) registration of medical
The Netherlands Cancer I nstitute Antoni van Leeuwenhoek Hospital Evolutionary optimization in (deformable) registration of medical images Ch. Siedschlag, J. Stroom, H. Blaauwgeers1,A. van Baardwijk2, H. Klomp, M. Wouters, K. Liesker1, R. van Pel, L. Boersma2, K. Gilhuijs 1: Onze Lieve Vrouwe Gasthuis, Amsterdam 2: Maastro Clinic, Maastricht Overview • Registration of images: necessary to get the most out of different modalities • Small or rigid deformations: affine transformations suffice • Bigger/elastic transformations: deformable registration • Automation with control loop -> EA • Sidetrack: EA in treatment optimization? 2 General optimization goal: optimize similarity • “Images” are two- or threedimensional arrays • For same modalit y: fitness function can be squared differences of grey values • For different modalities: corresponding points are colour-coded differently • Similarity measure of choice: mutual information 3 Mutual information Single image: grey value histogram conveys information Two images: correlated grey value histogram leads to measure of similarity (mutual information, based on minimalized entropy) 4 • Depending on the organ in question, deformations can be rigid (e.g. the head) or non-rigid (lung, breast etc…) • With rigid deformations: affine transformations are enough, not interesting from EA point of view (only 6 parameters in 3D, 9 if scaling is involved) • Nonrigid deformations are described by deformable regist ration: high number of parameters, nonlinear problem -> EA! 5 Deformable registration A deformation field is created by defining the displacement on a number of control points and finding an interpolation in between: n is typically 8-12 (in 2D) and 30 or more (in 3D). Optimization variables are the displacement vectors. Extended version: control points + displacement vectors Number of parameters: n*D(*2, if extended version) 6 A concrete study on lung cancer • CT (and PET) scans prior to surgery, detailed pathology analysis post surgery • Goal: match the modalities 7 The problem CT scan CT scan (masked lung lobe) PA photo match 3D rotate •Started in 2D •Has been extended to 3D, but not very successfully so far… 8 Example First phase: affine transformations Second phase: non-rigid deformations 9 Result of warping Has the global optimum been reached? Accuracy: 2-3 mm Is Mutual Informati on really behaving the way we expect it to behave? 10 The complete picture CT scan with tumor (red: GTV from CT, yellow: GTV from PET) and lung lobe (green) Warping Macroscopic and microscopic pathology results (dark blue: path. GTV, light blue: microscopic CTV) New margin recipes based on warped tumor cell distributions 11 The EA side • In collaboration with Ofer and Thomas: one particular algorithm (“DR2”, a (1,10) evolutionary strategy) behaved particularly well in laser pulse phase shaping • Contestor: Particle Swarm Optimization (PSO) 12 PSO algorithm • • • • • Modeled after a swarm of birds Ingredients: 20 “birds” with coordinate and velocity vectors in the search space Position update based on personal and social memory Parameter w, n1 and n2 are empirically determined, r1 and r2 are random numbers between 0 and 1 Advantage: extremely simple, can be coded in 5 minutes… Is it an EA at all? 13 Test runs with fixed control points DR2 PSO Fitness 8 CPs 12 CPs # of evaluations 14 Test runs with variable control points DR2 PSO Fitness 8 CPs 12 CPs # of evaluations 15 Test runs with variable control points and affine pre-matching (12 CPs) DR2 Fitness PSO # of evaluations PSO seems to be better , faster & more robust … 16 Explorative: fractionation optimization •Radiotherapy: treatment plan given by a set of doses D[i] at times t[i] •Normal procedure: 2 Gray (just a unit) per day, about 4 to 5 weeks •Recent advantages in beam shaping: concept should be looked at again. What’s the optimal treatment plan? Tumor healthy 17 Model: extended LQ Surviving cell fraction as function of dose D: 1 é ù S = expê- aD - b G(t R ) D 2 + s 2G(t S ) D 2 + H (TTot , Tk )(TTot - Tk ) / Tpd ú 2 ë û • Included: repair and resensitization times plus accelerated proliferation • For healthy tissue: only first two terms •Data available for tumor, early reacting normal tissue and late reacting normal tissue • Relevant quantity: biologically efficient dose BED = -1/a*ln(S) • Goal: keep normal tissue BEDs low , maximize tumor BED 18 H&N tumors (fast proliferating, large a/b), 40 Gy total delivered in 20 doses: Simple attempt: restricted to one delivery every 24 hrs Dose in normal tissue: 40 percent of tumor dos e 19 H&N tumors (fast proliferating, large a/b), 40 Gy total delivered in 20 doses: Unrestricted delivery times Dose in normal tissue: 40 percent of tumor dos e 20 H&N tumors (fast proliferating, large a/b), 40 Gy total delivered in 20 doses: Unrestricted delivery times Dose in normal tissue: 20 percent of tumor dos e Overall: relatively small effects for H&N tumors ! 21 Prostate tumor (slowly proliferating, small a/b), 40 Gy in 20 doses Dose in normal tissue: 20 % Restricted: one fraction per day Daily delivery: skew dose distribution can already more than double the tumor BED (due to the s mall a/b ratio) 22 Prostate tumor (slowly proliferating, small a/b), 40 Gy in 20 doses Dose in normal tissue: 20 % Unrestricted delivery times Huge gains! Much longer Might intervals be are optimal (15-20 days) something worth investigating… 23 Summary • Challenging optimization problems in deformable regist ration • PSO seems to be at least as good as ES3 • Step to 3D: everything is even more difficult • General thought: EA optimization of treatment possible? Models, retrospective study analysis? 24