imrt software design goals - isye · hard constraint: dose to the convex roi < 45 gy. hard...

25
IMRT software design goals Joe Deasy, Beong Choi, and Dan Low Washington University in St. Louis [email protected]

Upload: others

Post on 25-Sep-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

IMRT software design goals

Joe Deasy, Beong Choi, and Dan Low

Washington University in St. Louis [email protected]

Page 2: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

IMRT software design goals

• Clinical relevance of parameters and plan model• Steerability of plan changes in next plan iteration

or at “tweaking phase”• Efficiency/speed of feasible solution generation

in a plan iteration• Dosimetrically faithful and delivery-time-efficient

leaf segmentation• Accuracy, likelihood of arriving at the global

optimimum to within a target %.

Page 3: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Research goals in IMRT treatment planning

• Improve the clinical relevance of the prescription method. That is, make the IMRT or 3DCRT treatment planning statement more relevant to the clinical objectives, stated in terms of outcome goals (TCP or NTCP) or dosimetric goals (dose limits, dose volume constraints).

• Improve the responsiveness or “steerability” of IMRT treatment planning. That is, we want to improve the likelihood that a change to the planning algorithm input data will result in a predictable change in the resulting dose distribution.

Page 4: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Solve an optimization problem

Review: is this the best plan possible? & is it clinically acceptable? If “no” to either, change input and re-run.

Input algorithm parameters:Hard constraints, objective function weights. Solve a series of optimization

problems which add the next lower priority goal at each iteration. Higher priority goals are “constraints” in lower priority iterations.

Review: was the prescription statement appropriate? & is the resulting plan clinically acceptable? If “no” to either, change input and re-run.

Input prioritized prescription planning objectives

Current paradigmPrioritized prescription

optimization

Page 5: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Prioritized treatment goals• Prioritization of the prescription goals

– avoids tradeoffs among objectives which are difficult to control and sometimes clinically undesirable

– avoids fixing hard constraints to be more restrictive than necessary

– allows for more factors to be included in the prescription goals without degrading the most important goals.

• An optimization engine for prioritized prescription goals has been designed and implemented (PriOpt).Goals are incorporated iteratively. The planning engine is applicable to both IMRT and non-IMRT optimization.

Page 6: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Objectives• Improve the clinical relevance of the optimization input

data specification. That is, make the IMRT or 3DCRT treatment planning statement more relevant to clinical objectives, stated in numerical terms of outcome goals(TCP or NTCP) or dosimetric goals (dose limits, dose volume constraints).

• Improve the responsiveness or “steerability” of optimized treatment planning. That is, we want to improve the likelihood that a change to the planning algorithm input data will result in a predictable change in the resulting dose distribution.

Page 7: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Method (overview)• Prioritize the prescription goals. Prioritization

avoids tradeoffs among soft-constraints which are difficult to control and sometimes clinically undesirable. Prioritization also avoids suboptimally fixing hard constraint limitswhich could happen if multiple goals are incorporated up-front as hard constraints.

• We designed and implemented an optimization engine for prioritized prescription goals. As discussed below, goals are incorporated iteratively, in the order of their prescription-stated priority. The planning engine is applicable to both IMRT and non-IMRT optimization.

Page 8: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

MotivationWithout prioritized input prescriptions, there is no way to enforce such simple prescription statements as: – Primary goal: hold the spinal cord dose to < 45

Gy, – Secondary goal: then increase the minimum PTV

dose as much as possible, – Tertiary goal: then increase cell kill to the PTV as

much as possible.

Why? Because there is no a priori way to know the best minimum PTV dose when the maximum dose to the spinal cord is 45 Gy.

Page 9: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

2-D Test problem

Page 10: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Hard Constraint: Dmax to the convex ROI < 45 Gy.

Candidate goal: Dmin to the concave ROI > 80 Gy.

Base: Minimize quadratic sum of doses to non-ROI regions.

6 MV PBs, 2 cm wide, 0.4 cm voxels, 7 Ports

Page 11: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Hard Constraint: Dose to the convex ROI < 45 Gy.

Hard Constraint: Dose to the concave ROI > 52 Gy.

Candidate goal: Dmax to the concave ROI < 90 Gy.

Base: Minimize quadratic sum of doses to non ROI regions.

Page 12: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Hard Constraint: Dose to the convex ROI < 45 Gy.

Hard Constraint: Dose to the concave ROI > 52 Gy.

Hard Constraint: Dose to the concave ROI < 90 Gy.

Candidate goal: Minimize mean cell survival in the concave ROI.

Base: Minimize quadratic sum of doses to non-ROI regions.

Page 13: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Hard Constraint: Dose to the convex ROI < 45 Gy.

Hard Constraint: Dose to the concave ROI > 52 Gy.

Hard Constraint: Dose to the concave ROI < 90 Gy.

Hard Constraint: cell survival < 2.1 x 10e-8 in the concave ROI.

Candidate goal: Maximize the mean dose to the concave ROI.

Base: Minimize quadratic sum of doses to non-ROI regions.

Page 14: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

“Gee, isn’t this just pre-emptive goal programming?”

Page 15: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Prioritized optimization conclusions

• Prioritized prescription optimization introduces novel flexibility into the prescription statement.

• We hypothesize that: prioritized goal prescriptions allows for more clinically relevant statements of the treatment planning problem, including acceptable tradeoffs, compared to current (non-hierarchical) prescription statements methods.

• A more clinically relevant problem statement is likely to lead to fewer “back to the computer” plan iterations.

Page 16: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

The “tweaking phase”

• May need a different user interface than the initial algorithm inputs (Pollock, Henderson)

• Initial phase planning phase could be with simplified functions (Lee).

Page 17: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Isolated tradeoffs: a key missing capability of IMRT planning

systems (1/2)• What if the target cold spot is too large?

Or a normal tissue hot spot is too hot?– Where will we sacrifice/tradeoff dose to make

the desired improvement?– How can we control the tradeoff?

• The system must be capable of constraining doses to other normal structures do they don’t degrade as the tradeoff is made.

Page 18: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Isolated tradeoffs: a key missing capability of IMRT planning

systems (2/2)• This can be accomplished with

constrained optimization algorithms• Will be key to make plan modifications

predictable: “What will happen if I…”• In turn isolated tradeoffs and the resulting

predictability will be key to allow users to learn from experience, thereby decreasing trial-and-error in planning.

Page 19: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Dose-volume constraints imply multiple-local minima

Longitudinal axis

Suppose the green critical structure has a volume effect represented by dose-volume constraints

1 2 3 4 5 6

Where do we put the hottest regions/where do we put the coldest regions?

Page 20: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Plan-then-segment?

We are pursuing a plan-then-segment approach (commonly used now) instead of

a plan-with-segments (Hyperion) approach.

Page 21: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Why use plan-then-segment?

• It can look for the best plan without assumptions regarding segment shapes

• It will probably be faster than modifying segments

• The amount of degradation due to segmentation could be controlled (i.e. add more segments to get within 2% dose degradation everywhere)

Page 22: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Why use plan-then-segment?

• Current segmenters only try to match the fluence profile for each beam, we propose matching the dose distribution, which is a much more relaxed goal.

• In collaboration with Eva Lee of Georgia Tech

Page 23: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

IMRT algorithm tradeoffs

Model complexity/intractabilityClinical relevanceof model and parameters

Model iteration compute time

Number of planningiterations needed

Max desired operator time

Page 24: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

Parallel clusters

• To show multiple feasible solutions with different plan parameters (sensitivity analysis).

• To speed up optimization.• To speed up dose computation using

accurate methods (Monte Carlo)

Page 25: IMRT software design goals - ISyE · Hard Constraint: Dose to the convex ROI < 45 Gy. Hard Constraint: Dose to the concave ROI > 52 Gy. Hard Constraint: Dose to the concave

IMRT software design goals

• Clinical relevance of parameters and plan model• Steerability of plan changes in next plan iteration

or at “tweaking phase”• Efficiency/speed of feasible solution generation

in a plan iteration• Dosimetrically faithful and delivery-time-efficient

leaf segmentation• Accuracy, likelihood of arriving at the global

optimimum to within a target %.