Margins and margin recipes

Transcription

1 Margins and margin recipes Marcel van Herk On behalf of the image guidance group The Netherlands Cancer Institute Amsterdam, the Netherlands

2 Classic radiotherapy procedure Tattoo, align and scan patient Align patient on machine on tattoos and treat (many days) Draw target and plan treatment on RTP In principle this procedure should be accurate but

3 Things move: geometrical uncertainties Organ motion: largest error in prostate RT Baseline shift: largest error in lung RT In the past large safety margins had to be used

4 Example IGRT system: Elekta Synergy 1997: proposed by David Jaffray and John Wong 2004: prototype in clinical use at NKI 2005: Released for clinical use worldwide 6 at NKI, more than 500 worldwide Over scans made at NKI 200 GByte scans per week

5 With such a system, this is no longer needed to precisely irradiate a brain tumor

6 We can use this instead: focus on patient stability, but let computer position the patient with better than one mm precision Accuracy registration: 0.1 mm SD Accuracy table: 0.5 mm SD {x, y, z} Intra-fraction motion: 0.3 mm SD v Beek et al, in preparation

7 IGRT The good, the bad, and the ugly Good: IGRT gives unprecedented precision of hitting any clearly defined point in the body Bad: This precision may give us overconfidence in the total chain accuracy: tumors are rarely clear Ugly: we may have to find this out from our clinical mistakes

8 Nomenclature Gross error: mistakes, transcription errors, software faults: must be caught by QA Error: difference between planned value and its true value during treatment, however small Uncertainty: the fact that unpredictable errors occur quantified by standard deviations Variation: the fact that predictable or periodic errors occur

9 Mans et al, 2010 EPID dosimetry QA to catch gross errors: used for all curative patients at NKI Reconstructed EPID dose (VMAT case) EPID movie per frame cumulative Precision: within few %, enough to catch gross errors

10 Mans et al, 2010 Gross errors detected in NKI 0.4% of treatments show a gross error (>10% dose) 9 out of 17 errors would not have been detected pretreatment!!

11 What happens in the other 99.6%? There are many small unavoidable errors (mm size) in all steps of radiotherapy In some cases many of these small errors point in the same direction I.e., in some patients large (cm) errors occur(ed) This is not a fault, this is purely statistics What effect does this have on treatment? We do not really know!

12 Motion counts? Prostate trial data (1996) N=185 (42 risk+) N=168 (52 risk+) Risk+: initial full rectum, later diarrhea Heemsbergen et al, IJROBP 2007

13 The major uncertainties not solved by IGRT Target volume definition GTV consistency GTV accuracy CTV: microscopic spread Inadequacy of surrogate used for IGRT Motion that cannot be corrected Too fast Too complex

14 Delineation variation: CT versus CT + PET CT (T2N2) SD 7.5 mm CT + PET (T2N1) SD 3.5 mm Consistency is imperative to gather clinical evidence! Steenbakkers et al, IJROBP 2005

15 Effect of training and peer collaboration on teacher target volume definition students groups Material collected during ESTRO teaching course on target volume delineation

16 Glioma delineation variation (Beijing 2008) SD (mm) SD (mm) outliers removed Homework Margin (mm) Groups Validation Delineation uncertainty is a systematic error that should be incorporated in the margin Consistency is imperative to gather clinical evidence

17 Other remaining uncertainties Is the surrogate appropriate? 2.5 cm Motion of tumor boundary relative to bony anatomy

18 Are prostate markers perfect? Apex Base Sem. Vesicles +/-1 cm margin required Best: combine markers with low dose CBCT van der Wielen, IJROBP 2008 Smitsmans, IJROBP 2010

19 Intra-fraction motion: CBCT during VMAT

20 Intra-fraction motion: CBCT during VMAT This amount of intra-fraction motion is rare for lung SBRT

21 Error distributions Central limit theorem: the distribution of the sum of an increasing number of errors with arbitrary distribution will approach a Normal (Gaussian) distribution Large errors happen sometimes if all or most of the small sub-errors are in the same direction Normal distribution: 1,400 1,200 1, mean = 0 s.d. = 1 N = = 95%

22 Definitions (sloppy) CTV: Clinical Target Volume The region that needs to be treated (visible plus suspected tumor) PTV: Planning Target Volume The region that is given a high dose to allow for errors in the position of the CTV PTV margin: distance between CTV and PTV Don t use ITV for external beam! (SD adds quadratically)

23 Time-scales for errors Compare X planned with X actual X planned X actual = ε group + ε patient, group + ε fraction, patient, group + ε time, fraction, patient, group The appropriate average of each ε is zero X planned X actual = M g +/- σ g +/- σ p +/- σ f

24 The nomenclature hell Proposed to ICRU Bel et al. Literature M g Mean group error M Mean group error σ g Intra-group uncertainty Σ bias Inter-patient uncertainty (fraction) Systematic error σ p Intra-patient uncertainty σ Inter-fraction uncertainty (fraction) Random error σ f Intra-fraction uncertainty Intra-fraction uncertainty

25 van Herk et al, Sem Rad Onc 2004 Analysis of uncertainties Keep the measurement sign! patient 1 patient 2 patient 3 patient 4 fraction fraction fraction fraction Intrafraction Mean = 0.2 RMS of SD = σ f mean sd mean =M SD = Σ RMS = σ M = mean group error (equipment) Σ = standard deviation of the inter-patient error σ = standard deviation of the inter-fraction error { σf = standard deviation of the intra-fraction motion

26 Demonstration errors in RT Margin between CTV and PTV: 10 mm Errors: Setup error: 4 mm SD (x, y) Organ motion: 3 mm SD (x, y) 10 mm respiration Delineation error: optional

27 What is the effect of geometrical errors on the CTV dose? Treatment Random: execution Breathing, (random) intrafraction errors motion, blur the IGRT dose inaccuracy distribution CTV Systematic: Preparation delineation, (systematic) intrafraction errors shift motion, the dose IGRT distribution inaccuracy CTV dose

28 Analysis of CTV dose probability Blur planned dose distribution with all execution (random) errors to estimate the cumulative dose distribution For a given dose level: Find region of space where the cumulative dose exceeds the given level Compute probability that the CTV is in this region

29 x Computation of the dose probability for a small CTV in 1D 95% In the cumulative (blurred) dose, find where the dose > 95% x average CTV position 98%..and compute the probability that the average CTV position is in this area

30 What should the margin be? mm 6 mm 9 mm 0 mm minimum CTV Dose (%) Typical prostate uncertainties with bone-based setup verification

31 Simplified PTV margin recipe for dose - probability To cover the CTV for 90% of the patients with the 95% isodose (analytical solution) : PTV margin = 2.5 Σ σ Σ = quadratic sum of SD of all preparation (systematic) errors σ = quadratic sum of SD of all execution (random) errors (van Herk et al, IJROBP 47: , 2000) *For a big CTV with smooth shape, penumbra 5 mm

32 2.5Σ + 0.7σ is a simplification Dose gradients ( penumbra = σ p ) very shallow in lung smaller margins for random errors M = Σ ( σ p + σ ) 1.64σ p Number of fractions is small in hypofractionation Residual mean of random error gives systematic error Beam on time long respiration causes dose blurring If dose prescription is at 80% instead of 95%: M = Σ ( σ p + σ ) 0.84σ p (van Herk et al, IJROBP 47: , 2000)

33 Practical examples

34 Prostate: 2.5 Σ σ all in cm systematic errors squared random errors squared delineation Rasch et al, Sem. RO 2005 organ motion van Herk et al, IJROBP 1995 setup error Bel et al,ijrobp 1995 intrafraction motion total error times 2.5 times 0.7 error margin total error margin 1.27

35 Prostate: 2.5 Σ σ Now add IGRT all in cm systematic errors squared random errors squared delineation Rasch et al, Sem. RO 2005 organ motion van Herk et al, IJROBP 1995 setup error Bel et al,ijrobp 1995 intrafraction motion total error times 2.5 times 0.7 error margin total error margin 0.70 Engels et al (Brussels, 2010) found 50% recurrences using 3 mm margin with marker IGRT

36 CNS: single fraction IGRT for brain metastasis all in cm systematic errors squared random errors squared delineation organ motion setup error intrafraction motion total error times 2.5 times 0.7 error margin total error margin 0.30 Tightest margin achievable in EBRT ever due to very clear outline on MRI

37 Planning target volume concepts Convention Free-breathing CT scan Internal Target Volume exhale Mid- Ventilation /Position Timeaveraged mean position Crap Too large } Margin? Motion GTV/ITV CTV PTV

38 Image selection approaches to derive representative 3D data Vector distance to mean position (cm) 4D CT Exhale (for gating) Mid-ventilation

39 Very clear lung tumor: classic RT all in cm systematic errors squared random errors squared delineation organ motion setup error Intra-fraction motion 0 0 respiration motion (0.33A) total error times 2.5 difficult equation (almost times 0.7) error margin total error margin 1.47 Using conventional fractionation, prescription at 95% isodose line in lung

40 Very clear lung tumor: IGRT hypo all in cm systematic errors squared random errors squared delineation organ motion setup error 0 0 Intra-fraction motion respiration motion (0.33A) total error times 2.5 difficult equation non-linear error margin total error margin 0.89 Using hypo-fractionation, prescription at 80% isodose line in lung

41 Planned dose distribution: hypofractionated lung treatment 3x18 Gy

42 Realized dose distribution with daily IGRT on tumor (no gating) 2 cm 9 mm margin is adequate even with 2 cm intrafraction motion

43 But what about the CTV? By definition disease between the GTV and the CTV cannot be detected Instead, the CTV is defined by means of margin expansion of the GTV and/or anatomical boundaries Very little is known of margins in relation to the CTV Very little clinical / pathology data Models to be developed

44 Hard data: microscopic extensions in lung cancer % cases with extensions N=32 Deformation corrected distance from GTV [mm] 30% patients with low grade tumors (now treated with SBRT with few mm margins), have spread at 15 mm distance 100% 50% 25% Having dose there may be essential! Slide courtesy of Gilhuijs and Stroom, NKI

45 Is dose outside the prostate related with outcome? detect disease spread in historical data prostate Mapping of planned dose cubes to standard patient Dose differences due to: - randomization - anatomy - technique

46 Estimate pattern of spread from response to incidental dose in clinical trial data (high risk prostate patients) Average dose no failures average dose failures - = 7 Gy p = 0.02 PSA controls PSA failures 100% 80% median 60% 40% Witte et al, IJROBP2009; Chen et al, ICCR % 0% < median (53.1 Gy) p = Treatment group IV, Hospital A (n=67) Y

47 Conclusions We defined a margin recipe based on a given probability of covering the CTV with a given isodose line of the cumulative dose The margin with IGRT is dominated by delineation uncertainties Margins for random uncertainties and respiratory motion in lung can be very small because of the shallow dose falloff in the original plans

48 Conclusions In spite of IGRT there are still uncertainties that need to be covered by safety margins Important uncertainties relate to imaging and biology that are not corrected by IGRT Even though PTV margins are designed to cover geometrical uncertainties, they also cover microscopic disease Reducing margins after introducing IGRT may therefore lead to poorer outcome and should be done with utmost care (especially in higher stage disease)

49 Us Modern radiotherapy

50

51 Clinical motion estimation (rigid body registration) 4D CT Mid V Roughly paint mask % 0% Tumor trajectory Redundant registration (10 x 9 times) Entire procedure takes about 1 minute: Estimates trajectory, automatically removes outliers Nijkamp et al., ICCR 2007

52 Is full 4D planning necessary? 10% Tumor peak-peak amplitudes range from 1.1 to 3.6 cm relative difference mean dose GTV 5% 0% -5% -10% Conventional On-line Hypo } } mean = -0.9% patient The blurred mid-ventilation dose is almost the same as a full 4D dose calculation Mexner et al, ESTRO2007; submitted

53 Mid-position CT: in research 4DCT Deformable registration 4D DVF Deform 4DCT to local mean pos. Average frames Mid-position CT Wolthaus et al, Med Phys 2008 (in press)

54 Margins in lung hypo (3 x 18 Gy) Systematic Random Delineation 2 mm SD - Registration/couch shift 1.5 mm SD 1.5 mm SD Intra-fraction motion 1.5 mm SD 1.5 mm SD Total 3 mm SD 2.2 mm SD Margin A=10 mm 7 mm + 0 mm Margin A=20 mm 7 mm + 2 mm M Σ ( σ p + σ ) 0.84σ p = σp 7.8 mm Ensures 80% isodose encompasses GTV 90% of time in lung

55 Lung Treatment Baseline (off-line EPID, normal fractionation) Advanced (SBRT with on-line soft tissue based IGRT) Lung Systematic error Random error Systematic error Random error Delineation Baseline motion Setup error Intrafraction Respiration Total error Margin Total margin

56 Margins (cranio-caudal) GTV to PTV: classical delineation Steenbakkers et al, 2005 respiration x A (1 cm); van Herk et al., 2004 baseline shift Sonke et al, 2007 setup error NO EPID protocol total error root of sum square times 2.5 times 0.7 Simplification! error margin factors for random and systematic (2.5 and 0.7) 0.19 note: margin for SBRT total error margin 2.18 sum GTV-CTV margin 0.00 total margin 2.18 sum Clinical used margin: 1.5 cm cranio-caudal, 1 cm other directions 56

57 Margins (cranio-caudal) GTV to PTV: add technology delineation Steenbakkers et al, 2005 respiration x A (1 cm); van Herk et al., 2004 baseline shift Sonke et al, 2007 setup error NO EPID protocol total error root of sum square times 2.5 times 0.7 Simplification! error margin factors for random and systematic (2.5 and 0.7) 0.19 note: margin for SBRT total error margin 2.18 sum GTV-CTV margin 0.00 total margin 2.18 sum Clinical used margin: 1.5 cm cranio-caudal, 1 cm other directions 57

58 Uncertainty management: Conventional IMRT planning with margin CTV M = 2.5Σ+0.7σ PTV 90% prob. of D 95% D prescribed in CTV OAR Inverse optimization Objective functions Poisson cell kill, EUD, DVH points,... Dose distribution

59 Uncertainty management: Probabilistic biological IMRT planning without margin CTV no PTV margin! Maximum TCP for given OAR NTCP OAR Inverse optimization Objective functions with simulated errors TCP, NTCP Σ, σ Dose distribution

60 87 Gy 80 Gy 74 Gy 65 Gy 39 Gy

61 Points to add Effect of geometric uncertainty on TCP (1mm = x%) Vs effect of dosimetric unc on TCP (1% = 1%) Kill ITV Kill gating Volume of underdosage effect Read mans paper

62 Radiotherapy in the past (25 years ago)

63 Advanced imaging: 4D PET/CT Shows correct tumor shape and its components Shows range of respiratory motion Allows optimal tumor targeting taking motion into account

64 Steenbakkers et al, IJROBP 2005 Main problem now: target definition - 11 observers from 5 institutions, 22 patients - newly developed delineation software - delineation on CT + (one year later) CT+PET

65 Consistency is good Our clinical goal is to irradiate a target volume this volume should be well defined This does not mean that the well-defined target volume is correct! However, if all physicians (e.g., in a trial) agree on target volume, clinical experience allows us to learn whether it is right or not The alternative is pathology validation but this can only be done in surgery patients

66 Pathology validation (NKI/ MAASTRO/OLVG) Gilhuijs, Stroom and Boersma 34 surgical lung cancer patients CT 18 F-FDG PET Macroscopical Pre-op : Post-op : Analysis : CT, PET Pathology 1361 slides Microscopical

67 Accuracy CT/PET for gross tumor GTV PET: 42% Max SUV contours Effective diameter GTV CT (mm) Effective diameter GTV PET (mm) Effective diameter GTV pathology (mm) Effective diameter GTV pathology (mm) Slide courtesy of Gilhuijs and Stroom, NKI

68 Was this patient perfectly aligned after shifting the couch with IGRT?

69 Multiple clipbox registration Conventional single ROI registration (used for global patient setup) multiple ROI (mroi) registration (local misalignments) purple = planning CT green = cone beam CT Clinical in NKI, not commercially available

70 Pitfalls of IGRT Overconfidence in precision of delineation we use 5 mm margin for the clearest of tumors Intra-fraction motion generally requires only small margins Uncorrectable errors such as deformations and large rotations Anatomical changes warrant adaptation but for which patients and when? Too much focus on anatomy, not on treatment

71 Repetitive 4D CT: treatment response

72 Differential motion No couch correction can solve this problem

73 0.4 cgy 3 cgy Seeds allow low dose imaging 0.35 mm visicoils Seeds and soft tissue (seminal vesicles) visualized in low dose 1 minute scan

74 Hypofractionated lung treatment (3 x 18 Gy) 100x real speed Danger case: target close to spinal cord

75 SBRT soft tissue guidance Alignment to room lasers 4D-CBCT Registration 4D-CBCT Planning CT Bone Tumour Couch shift 4D-CBCT for verification First arc + 4D-CBCT Second arc + 4D-CBCT

76 Image respiration on cone beam CT? 2 x real time speed 3D reconstruction

77 Software demo of 4D lung SBRT about 200 patients done this way in NKI solution commercial since march 2009

78 Scattered Fluence + kv only MV scatter only

79 kv only

80 kv only vs kv+mv

81 kv only vs kv+mv-<mv>

82

83 Registration protocols in XVI4.5 Protocols: Clipbox only Mask only Dual registration Registration tools (for all workflows) Bone Seed Grey value 4D Grey value Workflow Registration Correction Overview

84 Incorrect deformable registration Prior After: looks OK Only edges are correc

85 Simulation: effect of microscopic spread on margin requirement Tumor with microscopic spread Iso-TCP Decreasing TCP (5% steps) Cell ratio GTV/Shell Solid tumor Margin 10 Microscopic spread is partly covered by the PTV margin and beam penumbra

86 Planning target concepts Convention Free-breathing CT scan Internal Target Volume exhale Mid- Ventilation Timeaveraged mean position Bad Too large } Margin? Motion GTV/ITV CTV PTV

87 Modes of Tumor Regression elastic erosion

88 Opportunities of IGRT mm precision Soft-tissue guidance Visualize normal structures Detect anatomical changes Novel correction protocols Novel fractionation schemes Monitor during treatment (VMAT) Daily 4D imaging Detailed knowledge of delivered dose

89

90 4D CT (PET): less artifacts + motion data Allows determination of correct shape, SUV, mean position and trajectory of tumor Fused 4DCT and 4DPET: Wolthaus et al, PMB 2005

91 What to do with this data? Full 4D planning and plan optimization: Not in this talk Motion estimation Image registration Derive representative 3D CT from 4D data Image processing Image selection Motion management Gating/tracking Motion inclusive planning

92 Image processing approaches to derive representative 3D data 4D CT Mean ( Slow CT ) Max ( MIP ) Free breathing-like Low resolution Density correct ITV-like Density over-estimated

93 Mid-ventilation is very simple (used clinically on hundreds of patients) Mid-ventilation CT 4D CT Eliminates systematic error due to imaging (except hysteresis) Geometrically and dosimetric very close to full 4D plan! Wolthaus et al, IJROBP 2006; Nijkamp et al ICCR 2007

94 Mid-position CT: deform all anatomy to its mean position and average over all frames Mid-ventilation image Mid-position image Reduces noise and artifacts Wolthaus et al, Med Phys 2008 (in press)

95 mean correction strategy: correcting the average error Possible correction strategies: Mean corrections Minimizarion of maximum error

96 Adaptive replanning on average anatomy daily CBCTs deformation vector fields N Planning CT Average anatomy systematic deformations Kranen et al, ESTRO2009

97 Inter-observer variation in delineation LOCAL SD (mm) > 15 SD is 2 mm at best: this is what we use for our hypo patients Steenbakkers et al, IJROBP 2005

98 No MV-Beam With MV- Beam Concurrent VMAT CBCT acquisition

99 Error distributions & central limit theorem Average of N uniform distributed numbers This works for any distribution

100 Soren Bentzen: Accuracy is quality not quantity priority of resources, assymptotic diminishing returns cost benefit level of QA for trials precision reproducibikity Q: rmse = s^2 + bias^2?; P (x>x) hangt niet af van RMSE Q: do we know the target, even the dose! Gamma DR for indentical mice about 4 (H.Suit), is this true if you irradiate the tumor only? Double trouble: Dose unc affects total dose & dose/fraction: GammaN > GammaD variation in dose delivery do not cancel when de2/dd2 not zero for 5% tcp loss: precision between 3 and 5% needed 1% dose accuracy: about 1% TCP, 1 mm about 5% loss importance of accuracy dose/geometry depends on your baseline

101 Ellen Yorke look at mellin 2011 pigs 1 fx d50 20 GY

102 Jeraj Dark ages bio im is qual imaging even in wisc did not pass national QA bowen et al NMB: tracer retention mechanism Q: tracer transfer function depends on goal due to nonuniformity! ph ( ) and other pars strongly affects Cu-ATSM response FLT short term gives perfusion mostly not prolif Q: can we fix shift effect on PET images

103 Demonstration margins in lung Margin between CTV and PTV: 7 mm Errors: Delineation error: 2 mm SD Registration error 2 mm SD Intra-fraction motion 2 mm SD Organ motion mm respiration