A new optimization engine was introduced in the Eclipse TPS (Varian Medical Systems, Palo Alto, USA) in the release 13.5. This is made of three main components: a model building and training engine; a model-based dose-volume histogram (DVH) and automated constraints prediction tool; a new VMAT and IMRT optimization algorithm to manage the above.
The sequence of main steps necessary to generate a model is as follows:

Selection of a set of training plans. No specific requirements are mandatory to be candidate for training. The strategy adopted was to create a “universal” model for HCC and the only requirement was to select plans accepted for clinical treatment.

Association of these plans to a model layout where target and organs at risk ontology, dose prescriptions and some descriptive elements can be defined.

Definition of the type of constraints to be generated per each structure (points vs. lines, priorities, user defined vs. fully automated).

Dosimetric and geometrical data extraction from the patient database to the model engine.

Model training. Based on principal component (PC) analysis methods [2 (link),3 (link),15 ,16 (link)].

Model publication and validation.

Figure 1 shows a schematic representation of the model determination steps and of the PC method applied to DVH. The assumption is that any DVH can be represented as a combination of the average DVH over a population plus a sum of some weighted PC and a residual. The first PC is determined by maximizing the variance of the training set it can explain; any consecutive PC is chosen so that the residual variance is further accounted for. The features used to build the model primarily include geometric characteristic of the various structures, as well as their mutual position and their relationships with the treatment fields. These are modeled by constructing a Geometry-Based Expected Dose (GED) which evaluates the distance between each structure and the target surface by means of the amount of dose that each target contributes to an organ for the current field geometry. The final prediction model is built as a combination of the PC and regression techniques for the in-field region of any OAR and a mean and standard deviation model on the DVH fo the other OAR regions. The PC is applied to the GED and DVH to find the main component scores.

A schematic representation of the model determination steps and of the PC method applied to DVH.

A trained model, once made “public”, can be used to perform predictive estimation of the DVHs for any given new test case and, from these, to determine the planning constraints. The DVH prediction workflow is described as:

Selection of a knowledge-based model

Matching of structure names if the ontology mapping is not complete

Prediction of a range of possible DVH for each of the structures present in both the plan and the model.

Automatic generation of the dose-volume constraints based on the rules from the model configuration. With a fully automatic procedure, these are located below the lower limit of a prediction range generated from the most probable DVH curve by adding and subtracting a variation curve. This corresponds to 1 standard deviation for the out-of-field region. For the in-field region this is constructed by adding in quadrature the DVH PC multiplied by the standard error related to the model regression. Also the priorities are defined by the prediction engine and account for a basic balance between all possible trade-offs. All point constraints and priorities can be modified during optimization.

Figure 2 exemplifies the resulting model-based predictive objectives with the estimate range and automatic objectives (line objectives in this example).

Examples of the model-based predictive objectives with the estimate range and automatic objectives (line objectives in this example).

Free full text: Click here