A guide to using multi-criteria optimization (MCO) for IMRT planning in RayStation

Transcription

1 A guide to using multi-criteria optimization (MCO) for IMRT planning in RayStation By David Craft Massachusetts General Hospital, Department of Radiation Oncology Revised: August 30, 2011 Single Page Summary ('MCO Planning Cheat Sheet') Good MCO planning = a good choice of objectives and a good choice of constraints. Constraints are easier so we'll start with them: Use at least two constraints: 1) one to get some dose into the target 2) one to prevent too much dose to the patient. Here something like a maximum dose at 1.15* prescription to the entire patient makes sense, or a dose volume constraint. For additional constraints, only put constraints on that are an absolute must, meaning the plan would be flat out rejected by the physician if this constraint were violated. Don't put constraints for things you want but might have to live without. Objectives are used to navigate and steer dose in and out of structures, and to balance the tradeoffs inherent in radiotherapy planning. For every structure for which it is important to tweak the dose to in order to get the right plan, you should put one objective on. For OARs, the general recommendation is: use MAX EUD with dose level=0 and a=2. There are times when other functions are appropriate, but this will work for most OARs in most situations. This is discussed further in the main document. For the unclassified tissue (i.e. the entire patient voxel set) use a single Dose Falloff objective, which strives to conform the dose to the target(s). For the function parameters, use the prescription level for the high dose level, 1 cm (the default) for the distance, and 50% of the prescription level, or just 0, for the low dose level. For targets, use Min Dose with the dose level set at the prescription level and possibly Max Dose also at the prescription level for an additional slider that controls hot spots within the target. Alternately, use just the Uniform Dose objective to get one slider that simultaneously strives for both of these. Do not use all 3 at once since they are redundant. Rather than put every objective under the sun in there, just put the ones you think you need a lever for. More objectives slows things down. The rest of this document flushes these simplified instructions out quite a bit, explains their rationale, and indicates when one should deviate from them.

2 1. Introduction RaySearch Laboratories has produced a multi-criteria optimization (MCO) treatment planning utility for intensity modulated radiation therapy (IMRT). MCO aims to drastically shorten the planning time, but there are some key differences in how the planning is done with MCO versus a traditional IMRT optimization system. The purpose of this document is to explain how to use MCO effectively. MCO involves two main steps: setting up the problem (i.e. choosing your constraints and objectives, and then hitting the «Generate» button to do the database plan optimizations) and navigating the resulting plan space using the sliders. There will be one slider for each objective function defined. The more objectives there are the more time it will take to generate the database, so you will want to be somewhat careful about choosing not too many and not too few objectives. 2. Problem set up To use MCO effectively, you must differentiate between constraints and objectives. If I know two planning goals are in strong conflict with one another, I put them both as objectives. The main example of this is PTV coverage versus nearby critical organ sparing. Although the prescription from the physician might give DVH points for each of these, it may very well be that it is impossible to satisfy both organs simultaneously. Therefore, I put an appropriate objective for each organ and leave it until the navigation phase to explore this trade-off. 2.1 Constraints A vital feature in the MCO system is constraints. In optimization parlance, there are constraints and objectives. Constraints are mathematical relationships that have to hold for a solution to be considered feasible (constraints are sometimes also called hard constraints). For example, if the doctor requests that no voxel of the spinal cord receive dose above 45 Gy, and the doctor is not willing to sacrifice on this demand at all, this is a constraint. The danger in using constraints comes when a user specifies multiple constraints that are mutually incompatible. Despite this potential drawback, constraints are very useful in optimization. When using constraints in RayStation MCO, the user should make certain that there is no way the complete set of constraints forms an infeasible problem. A common mistake here would be the following scenario. Physician requests that all target doses >= 59.4 Gy and that the brain stem dose <=50 Gy, so the user puts these in as constraints. The problem is though that the brain stem and target are so close (or they overlap) and it is not possible to achieve both of these. While other treatment planning systems deal with that by finding a plan that is some compromise between these two constraints (i.e. the constraints are interpreted as soft constraints), RayStation MCO has been designed to interpret constraints exactly, and so one would not put these both as constraints. Instead, one could put one of these, or possibly put neither of them, and instead turn them both into objectives (discussed below). When the Generate button is hit, if constraints are not all satisfiable, the system will tell you so and stop. The following constraints are recommended:

3 An absolute upper bound on all the voxel doses. For example, if the prescription is 59.4 Gy, an upper bound of 1.15*59.4 might be appropriate (note although in this document I jump between Gy and cgy, in RayStation the dose is always in cgy). Some type of coverage constraint. At the minimum, a single dose volume (DV) constraint or minimum dose should be used for each target (we will assume a single target for the rest of the discussion). Assuming prescription, Rx, is 59.4, we might set V59.4 >= 90%, that is the volume of the target receiving the prescription or greater is at least 90%. Even if this is way worse than a physician would accept, it is OK, because we will take care of getting better coverage by using objectives. For these constraints, it is better to be too loose than be so tight that it is not possible to achieve it. So why have these constraints at all, why not just be very loose and drop the contraints all together? The reason is that when the multi-criteria optimizer is working on minimizing dose to the organs at risk, if it doesn't have at least one constraint on target coverage, the optimal solution will be to not dose the patient at all, and this is not a useful solution. Any constraint that cannot be violated (i.e. the doctor will reject a plan if such a constraint is violated). One should not interpret the set of constraints as sufficient for generating good plans. They are necessary, but not enough, and are only used to get us in the ballpark of good plans. The objectives are how we will create good treatment plans. 2.2 Objectives For organs at risk (OARs), we recommend Max EUD as the objective type. (Although RayStation allows DV-based objectives, these are not generally recommended because if they happen to be always satisfied then the Pareto surface will be degenerate in the sense that the slider associated with that DV objective will do nothing. And even if they are not always satisfied but can be satisfied, the optimizer will not try to push the dose down any further. So, even if the physician's prescription sheet has statements in the form of DV constraints, these should either be put in as constraints if they are strictly required and mutually satisfiable, or a Max EUD objective should be used and the DV goals should be sought after during the navigation step). There are two parameters that need to be specified for each EUD: dose level and exponent a. For dose level, I recommend always using 0 (until we find a good reason to use something different). For exponent a, I recommend 2. This is then the commonly used quadratic dose penalty which tried to minimize all OAR voxels doses but penalizes the hotter doses more. Max Dose is not recommended for OARS, even ones where maximum dose seems to be a good indicator of toxicity, because if the structure overlaps a PTV there is almost no control over the maximum dose, and even if there is no overlap, EUD with a=2 will focus on the high doses while still reducing all doses that it can. For the skin (i.e the unclassified tissue) I recommend using the Dose Falloff function. This brings down dose away from the target, and the farther away a voxel is from a target, the more emphasis there is on getting that dose down. Thus, it encourages conformal plans. This function has 3 parameters: high dose level, low dose level, and low dose distance. For these values I recommend the prescription dose for the high dose level, ½ the prescription dose for the low dose level, and 1 cm for the low dose distance. If there are multiple targets with different prescriptions, I use the lowest prescription value for the high dose level. For targets, I recommend minimum dose (as in maximize the minimum dose) as the objective type, with the precription level entered as the dose level. One can also have a max dose objective on a target which would be used for controlling the hotspots. Target Uniformity, which combines the min and the

4 max functions into one, is also useful if uniformity of dose is a strong requirement from the physician. But if you want separate control over min and max doses, better to use them individually. I typically have max dose as a constraint and maximize the minimum dose as an objective because target underdosage versus OAR sparing is the typical tradeoff physicians care about, more so than target uniformity versus OAR sparing. I do not recommend EUDs for targets. At least two objectives are needed, and if only two are used, the user should verify that these objectives are in conflict with one another. Regarding how many objectives to use, the user should have as many as deemed important players in coming up with a good plan. If an OAR is contoured but is way out of field and will not receive much dose anyhow, there is probably no reason to have it as an objective. In this case I either ignore it completely or put on the physician's DVH points for it as constraints. 3. Database Generation Once the objectives and constraints are defined, the user will generate the database. There are two options for this: Standard and Enhanced. Let N denote the number of objectives defined. For Standard, N+1 plans will be generated: one for each objective independently, and one balanced plan. For Enhanced, 2N plans will be generated. It remains to be seen if Enhanced is clinically advantageous. My current recommendation is to use Standard if you are in a rush, but if you are heading off to lunch anyhow, use Enhanced. 4. Navigating OAR sliders are pulled left to decrease the dose to an OAR. Min dose target sliders are oriented the other way, so dragging them right improves the minimum dose (the consistency is: leftward slides decrease dose and rightward slides increase dose). Locks, the check boxes beside each slider, are used to generally keep that particular objective from getting any worse than it is when the lock is invoked. But one should not be surprised to see some slippage when locks are in place, and this is due to mathemtical approximations that are being made under the hood. Locks should be seen as helping you steer the navigation and not as exact clamps on the current DVH. One can also expect to sometimes see all of the OAR sliders improve when one of them is improved. This is because in the current implementation, the user is permitted to explore the entire set of feasible plans that arise from combining the database plans, not just the Pareto optimal set. This allows for a richer set of plans to view. 5. Making the plan deliverable After navigation, which happens in idealized fluence map space, the plans need to be segmented for delivery. For plans where the navigated-to dose distribution is not too complicated (i.e. the iso dose lines near the target are nearly convex) then the default deliverable parameters (50 total segments, 40 iterations, etc.) should be fine. I recommend increasing the maximum number of segments for harder cases that have for example horseshoe-shaped dose distributions. The other relavant parameter that you might want to play with is Target Coverage Priority. The system will try to match the DVHs that you navigated to when you hit «Start». But, matching the target DVH is in conflict with matching the OAR DVHs. This parameter allows some control over that balance. The default 10 is usually fine, but if your deliverable plan sees too much target slippage, and increasing the number of segments doesn't help, you might try to increase Target Coverage Priority to 20 or higher. Final dose computation is required after making the plan deliverable. This does not change the segmentation, just performs a better dose calculation.

5 5. Examples We present three cases to illustrate the above ideas. We start with actual planning goals as given by the physicians and show how to turn this into an MCO formulation. A general comment is that it is preferable to set up a case with the main tradeoffs as objectives instead of including every possible OAR and target, which will slow down the database generation time considerably. These formulations are a bit on the heavy side because I wanted to indicate how you might set up a problem without too much experience in that particular disease site. With experience, one might use smaller formulations for each of these examples. 5.1 Pancreas FROM PHYSICIAN Panc IMRT PTV to 50.4 Gy Spinal cord max 43 R and L kidney V18<20, V10<50 Liver V30<30, mean<24 Stomach mean<14, V40<10 Formulation As is typically the case, these prescriptions and constraints from the physician are desires, but there is no certainty that they can all be met. Therefore, we turn this into the following MCO formulation: Objectives: 1) PTV min dose 5040 cgy 2) R kidney EUD a=2, dose level=0 3) L kidney EUD a=2, dose level=0 4) Liver EUD a=2, dose level=0 5) Stomach EUD a=2, dose level=0 6) Cord max, dose level 0 7) unclassified tissue: Dose Falloff: high dose level 5040, low dose level 2520, distance 1 cm. Constraints: Min dose to PTV: 0.9*50.4 = Gy Dose at volume 90% for PTV = 50.4 Gy. Global max dose: 1.12*50.4 = 56.5 (i.e. allow hot spots up to 12% above prescription)

6 During navigation one finds that all of the OAR DV constraints can be achieved, but this is not the case for all pancreas patients, and so although it would have been OK in this case to put some as constraints (for example, the cord and the kidneys) and then explore a lower dimensional tradeoff (PTV coverage, liver, stomach), this will not generally work and so is not advised without experience in the particular treatment site.

7 5.2 Glioblastoma multiforma (brain) FROM PHYSICIAN Prescription dose; 60 Gy Technique: IMRT. Normal tissue constraints: Brainstem max 54 Gy, surface to 60 Gy, try to keep < 2% Optic nerves/chiasm: 54 Gy Eyes: 40 Gy Lens: 8 Gy Formulation As seen in the figure, the right optic structures are far enough from the target that they do not need to be included in the tradeoff, but the right lens is included as a constraint. It would be fine to put the right eye as a constraint too. Left eye is included as an objective. Left lens could be too, but the doses to those structures correlate so those sliders would move together. In such cases it is better to just put the main structure in. Therefore we use the following 6D tradeoff. Objectives: 1) CTV min dose 60 Gy 2) brain stem EUD a=2, dose level=0 3) left optic nerve EUD a=2, dose level=0 4) chiasm EUD a=2, dose level=0 5) left eye EUD a=2, dose level=0 6) unclassified tissue: Dose Falloff: high dose level 60, low dose level 30, distance 1 cm. Constraints: Min dose to =CTV: 0.9*60 = 54 Gy Dose at volume 90% for CTV = 60 Gy. Right lens max dose: 8 Gy Global max dose: 1.10*60 = 66 (i.e. allow hot spots up to 10% above prescription)

8

9 5.3 Anal cancer RTOG guidelines give a set of DVH constraints for targets and OARS (See pages 9-12). After experimenting with these constraints on a few different cases, we observed that they were all achievable simultaneously. Converstations with the physician revealed that the main tradeoff for this site is between small bowel sparing and the coverage of the PTVs. This led to the following formulation (all DVH constraints are directly from the RTOG specification). Formulation (at the time this image was made, Dose Falloff was termed Target Conformance)

10