Image Guided Adaptive Radiotherapy Model Di Yan, D.Sc. Radiation Oncology William Beaumont Hospital IGART Process: A treatment process which includes the individual treatment information, such as patient anatomical variation and dose in organs of interest assessed during the therapy course, in the treatment evaluation and planning optimization 4D Adaptive Planning Reference plan Treatment Dose Delivery Process Image (patient anatomy) & Machine Output Feedback Treatment Evaluation & Plan Modification Decision-Making Daily & Cumulative Dose Construction & Estimation IGART Model: Treatment Dose Construction Daily & Cumulative dose delivered to a organ subvolume v in organs of interest, VOIs, during the n fractions of treatment delivery Ti D (v ) = n ∑∫ i =1 t ∈T i v d ( x t ( v ) , Ω ( t ) , Φ ( t ) ), ∀ v ∈ VOIs v xt ( v ) − Subvolume displacement at the time t (Process Variable) Ω(t ) − Patient global matter distribution represented by a CT image obtained at the time t (Process Variable) Φ (t ) − Machine output at the time t (Control Parameter) IGART Model: Adaptive Management Optimal mapping (control law), determined before any treatment delivery day k < n, from the space of patient anatomical variation v (process variables), { x t , Ω ( t )}, to the space of therapy machine output (control parameter), Φ ( t ) , Φ * (t ) v = G ( x t ( v ) , Ω ( t ) | v ∈ VOIs ), n t ∈ U Ti i=k which optimizes the predefined objective function, v ⎛ F ⎜ ∫ d ( x t ( v ), Ω ( t ) , Φ ( t ) ) | ∀ v ∈ VOIs ; i = 1,... , k ⎞⎟ ⎠ ⎝ t∈Ti Lecture Outline 1. Description of patient anatomical variation process v { x t , Ω ( t )} 2. Treatment dose construction & estimation { Dˆ t } 3. Treatment evaluation & adaptive planning optimization { Φ ∗ (t) } Patient Anatomical Variation Process Organ displacement during the treatment delivery can be completely described using a Random Process, n ⎧v ⎫ v v ⎨ xt ( v ) = xr ( v ) + u ( t , v ) t ∈ T = U Ti ; v ∈ VOI ⎬ i =1 ⎩ ⎭ v Where subvolume displacement, u ( t , v ) , has a probability density v function (pdf) characterized by the mean, μ ( t , v ), (the systematic v displacement) & the standard deviation, σ(t, v) , (the random displacement). Patient Anatomical Variation Process 1. Stationary Process (i.e. daily setup induced variation) if ∀ v ∈ VOI , v v v u ( t , v ) ~ pdf ( μ ( v ) , σ ( v )) has the ‘time-invariant’ (constant) mean & standard deviation 2. Non-stationary Process (i.e dose response induced variation) if ∃ v ∈ VOI , v v v u ( t , v ) ~ pdf ( μ ( t , v ) , σ ( t , v )) has the ‘time-variant’ mean or standard deviation Process Parameter: Parametric Estimation Least Square Estimation Using {u( t1 ) , ⋅ ⋅⋅, u( tk ) } v Given an orthogonal base φ (t) = [ φ 1 ( t ) , φ 2 ( t ) , ⋅ ⋅⋅ , φ m (t) ∑ (u k μˆ( t ) = aˆ1 ⋅ φ1 ( t ) + aˆ2 ⋅ φ2 ( t ) + ⋅ ⋅ ⋅ + aˆm ⋅ φm ( t ) ; σˆ ( t ) = v ⎡ aˆ1 ⎤ ⎡φ ( t1 ) ⎤ ⎡ u( t1 ) ⎤ ⎢ aˆ ⎥ ⎢ 2 ⎥ = (Φ T Φ )−1 Φ T U , Φ = ⎢ M ⎥ ; U = ⎢ M ⎥ ⎢v ⎥ ⎥ ⎢ ⎢ M ⎥ ⎢φ ( t k ) ⎥ ⎢ u( t ) ⎥ ⎢ˆ ⎥ ⎣ k ⎦ ⎦ ⎣ a m ⎣ ⎦ ] i =1 ˆ ( ti ) ( ti ) − μ k −m ) 2 Parametric Estimation Residuals: μˆ ( t ) − μ ( t ) ≤ tα / 2 ,k − m ⋅ σ (t ) k ; σ (t ) ≤ k −m χ 2 1−α , k − m ⋅ σˆ ( t ) Example: v φ (t) = [ 1 , t , ⋅ ⋅⋅ , t m ] and if aˆ1 ≠ 0, aˆ2 ≈ 0 ,..., aˆm ≈ 0 k k u ( ti ) μˆ = aˆ1 = ∑ , i =1 k σˆ = 2 ˆ ( u − ) μ ( ) t ∑ i i =1 k −1 Example: Prostate Isocenter Displacement output 1st week 2nd Week 3rd Week 4th Week 5th Week 1.5 1 0.5 0 -0.5 -1 1 2 3 4 7 10 13 16 21 24 27 30 33 -1.5 2 3 ˆ ˆ ˆ ˆ ˆ μ ( t ) = a1 + a2 ⋅ t + a3 ⋅ t + a4 ⋅ t 36 39 42 45 Treatment (4D) Dose Construction & Estimation Dose summation of subvolume v in VOI with using CT samples {CT1, …, CTk}, (assuming identical machine output in each treatment delivery) D (v ) = n ∑∫ i =1 ( t∈Ti v d ( x t ( v ), Ω t , Φ r )⋅ dt n v v ≈ ⋅ d ( x t1 , Ω CT1 , Φ r ) + L + d ( x tk , Ω CTk , Φ r ) k ) Treatment Dose Construction & Estimation D (v ) = n ∑∫ v d ( x t ( v ) , Ω ( t ) , Φ ( t ) )⋅ dt ∫∫ v v v v 0 v TE ( x ' , Ω( t ) , Φ E ( x ' , t )) ⋅ AE ( x '− xt ( v ), Ω( t )) ⋅ dx ' dE dt n i =1 n = ∑∫ i =1 t ∈T i v t∈Ti E x '∈R v x' ⎛ v⎞ ⎛ liso ⎞ v μ 0 v ⎟⎟ ⋅ Φ E ( x ' , t ) ⋅ exp⎜ − ( TE = ( v ) E ⋅ ⎜⎜ ) E ⋅ ∫ ρ t (ξ )dξ ⎟ ⎜ ρ water ⎟ ρ t ( x ' ) ⎝ l xv ' ⎠ 0 ⎝ ⎠ v v AE = A( E , ρ t ⋅ l ( x ' − x t ( v ))) μ 2 Effective Density? Treatment (4D) Dose Estimation Using “The Central Value Theory of Integral”, there exists ‘a single representation of the global matter distribution’ (the mean density) such that the spatial dose distribution calculated using the ‘mean density’ is time-invariant. In this case, the summation dose can be estimated convolving the spatial dose calculated with the mean density to the estimated pdf, such as Dˆ k (v ) = ∫v x∈R v v ˆ v v vˆ vˆ Dp ( xr ( v ) + u , Ω r , Φ r ) ⋅ pdf (u , μ , σ ) ⋅ du 3 ⎛ vT v ≈ D p ( xr ) + ⎜⎜ μˆ ⋅ ∇ D p [ xv r , x r ⎝ 2 ∂ Dp ( xr 1 vˆ T vˆ ⋅σ ⋅ + μ] + v2 ∂x 2 v + μˆ ) vˆ ⎞⎟ ⋅σ ⎟ ⎠ Treatment (4D) Dose Estimation Pelvic & Abdominal Region ¤ The planning CT image can most likely be used as the single representation of the global matter distribution (if there is no large gas filling) H&N Region ¤ The planning CT can be applied if patient does not loss weight significantly Chest Region & Skin Surface ¤ Using CT image obtained at a single breathing phase in cumulative dose evaluation can cause as much as 6% discrepancy. However, the weighted sum of 4D images (the average image) can be applied as the single representation of the global matter distribution to calculate the time-invariant spatial dose distribution for dose convolution Example: Patient Respiratory Motion v v v D( x) = w1 ⋅ d ( x , ΩCT1 ) + L+ w10 ⋅ d ( x , ΩCT10 ) v v Δ = D( x , ΩCTi ) − D( x ) , 4 Dose Deviation% 3 2 1 0 0 1 2 3 4 5 6 7 8 9 10 -1 -2 -3 RepiratoryPhase v v ˆ ˆ Δ = D ( x ) − D ( x , Ω CT1 + ⋅⋅⋅ + CT10 ) = 10 0.4% -0.3% -0.19% Adaptive Management Application of adaptive management should not be limited by imaging modality & correction actions (online or offline) Patient variation and treatment dose must be considered in the adaptive planning modification Ideally, a continue optimal control from the space of patient anatomical variation (process parameters) to the space of therapy machine output (control parameters) should be applied ¤ However, it is either impossible or clinically impractical to perform such control mechanism Evaluation & Decision-making Decide if the on going treatment plan needs to be modified ¤ Treatment evaluation based on dose-volume factor, EUD, NTCP or TCP model has been well established, and can be applied to determine the on going treatment quality and make a decision if the planning modification needs to be performed Evaluation & Decision-making CT V Cumulative Dose 62.4 62.2 Planned Delivered EUD (Gy) 62 61.8 61.6 61.4 61.2 61 0 1 2 3 4 5 6 7 8 T reatment Fraction In Planning: TCP = 86% Tx Day: 85%, 84%, 83%, …………………………..…, 83% 9 10 Evaluation & Decision-making Brain Stem Cumulative Dose 38.5 38 37.5 Planned Delivered 37 36.5 36 0 1 2 3 4 5 6 7 8 9 Treatment Fraction In Planning: NTCP = 0.15% Tx Day: 0.21%, 0.18%, 0.22%, 0.22%, 1%, ……………… ., 1.2% 10 Adaptive Management: Patient-specific Target Margin Image sampling during the treatment course the individual systematic and random errors from the image registration (parametric approach) or ROI occupancy distribution (nonparametric approach) Estimate Correct the systematic error by adjusting patient position or beam aperture Construct the patient–specific CTV-to-PTV margin to compensate for the random error and the residuals Patient-specific Target Margin (Parametric Approach) 0.7 0.02 0.6 P r o b a b ilit y 0.5 0.4 0.3 0.2 0.012 0.008 0.004 0.1 0 -0.1 0.016 0 10 20 30 40 Time (second) u(t ) ; t ∈Ti 50 60 0 -0.1 0.1 0.3 0.5 Exhale --> Inhale pdf ( μ , σ ) 0.7 Patient-specific Target Margin (Parametric Approach) 2 D ( xref ) v ∂ 1 vT ⋅σ Δ = ⋅σ ⋅ v 2 2 ∂x Respiratory Motion: 2.0 cm Dose without motion Dose with motion Adaptive Management: 4D Inverse Planning Image sampling during the treatment course Deformable organ registration (parametric approach) or ROI occupancy distribution (nonparametric approach) Include organ motion information directly in the objective function for optimal plan search 4D Adaptive Inverse Planning (Parametric Approach) Offline: Determine the optimal beam intensity map, Φ∗, for the rest of treatment delivery (k+1 to n) with using the patient anatomical position and treatment dose observed during the k previous dose deliveries, ( Max F Dˆ k ( v , Φ Φ ) |v ∈ VOIs ) k v v v v vˆ vˆ ˆ Dk (v, Φ) = ∑di (v) + (n − k ) ⋅ ∫ 3 d ( xr ( v ) + u, Φ) ⋅ pdf (u, μ , σ ) ⋅du i =1 R Birkner M, Yan D, Alber M, Liang J, Nusslin F (2003) Adapting inverse planning to patient and organ geometrical variation: algorithm and implementation. Med Phys, 30(10):2822-2831 Dose & Occupancy Frequency(%) 4D Inverse Planning for Prostate Cancer 100 Rectal Density 80 60 40 20 0 31 Target Density 29 27 25 23 21 19 Anterior-Posterior Position (cm) 17 15 Conventional Inverse Planning (5 mm margin) 4D adaptive Inverse Planning CTV1 CTV2 Mandible Brain Stem Cord R Parotid L Parotid 4D Adaptive Inverse Planning (Nonparametric Approach) Offline: ( v v ˆ Max F Dk ( x , Φ ) | x ∈ VOIs Φ v ˆ Dk ( x , Φ ) = ∫v k UVOIi x∈ ) v v v d ( x , Φ ) ⋅ w( x ) ⋅dx i =1 Baum C, Alber M, Birkner, Nusslin F (2006) Robust treatment planning for intensity modulated radiotherapy of prostate cancer based on coverage probability. Radiotherapy & Oncology, 78:27-35 4D Adaptive Inverse Planning (Parametric Approach) Online: Determine the optimal beam intensity map, Φ∗, for the k+1th treatment delivery with using the instant patient anatomical position as well as all previous observations (anatomy and dose) during the k previous dose deliveries ( Max F Dˆ k ( v , Φ Dˆ k ( v , Φ ) = Φ ) | v ∈ VOIs ) k n ⋅ ( ∑ d i ( v ) + d k +1 ( xv ( v ), k + 1 i =1 Φ )) 90 80 80 70 70 60 60 100 90 Volume (%) 90 Volume (%) Day 2 Day 4 100 50 Day 5 100 50 90 40 40 30 30 80 80 70 70 Volume (%) 20 Volume (%) 20 60 10 10 50 0 1000 2000 3000 4000 5000 6000 7000 8000 0 9000 60 50 40 0 0 40 1000 2000 3000 4000 5000 6000 7000 8000 9000 Dose (cGy) Dose(cGy) 30 30 20 20 10 10 0 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 Dose (cGy) 0 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 Dose (cGy) Day 6 Day 1 100 100 90 90 Prostate+SV 80 80 70 Online Plan 70 Rectal Wall 60 Volume (%) Volume (%) Day 3 100 50 Online Adaptive Plan 60 50 40 40 30 30 20 20 10 10 0 0 0 1000 2000 3000 4000 5000 Dose (cGy) 6000 7000 8000 9000 0 1000 2000 3000 4000 5000 Dose (cGy) 6000 7000 8000 9000 Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 O n lin e P la n 100 Day_1 Day_2 80 Day_3 Volume (%) O n lin e A d a p t iv e P la n 100 volume (%) 80 60 Day_4 D ay_1 Day_5 D ay_2 40 Day_6 D ay_3 D ay_4 60 20 D ay_5 D ay_6 40 0 1000 20 2000 3000 4000 5000 D o s e (c G y ) 0 1000 2000 3000 4000 5000 d o s e (c G y ) 6000 7000 8000 6000 7000 8000 IGART Model: Summary ( ) MaxF Dˆ k ( v, Φ) | v ∈VOIs Φ Φ* Treatment Dose Delivery Patient Variation Process v ⎧ xt ( v ), Ω(t ) , Φ(t ) :⎫ ⎨ ⎬ ⎩ v ∈VOIs ; t ∈ Tk ⎭ Dp ( ) f Dˆ k (v) − Dp (v) | v ∈VOIs ≥ δ ? v ⎧d (xt (v), Ω(t) , Φ(t)) : ⎫ ˆ Dk (v) ~ ⎨ ⎬ ⎩v ∈VOIs; t ∈Ti , i = 1,..., k ⎭