Session Item

A bi-objective optimization model was recently introduced that entails direct optimization on dose volume indices (DVIs) specified in a clinical protocol. Optimizing this model gives a set of treatment plans that represents trade-offs between target coverage and organ sparing, which was shown to include clinically desirable plans for prostate HDR BT. We studied the direct extension of this approach to cervical cancer BT, by optimizing BT planning aims recommended by the EMBRACE-II protocol.

The objectives in the optimization model are the Least Coverage Index (LCI) and Least Sparing Index (LSI). Both indices are functions that return the DVI (pertaining to coverage or sparing, respectively) that is either most violated or least satisfied, making it the DVI that is currently iteratively being optimized. The used optimization algorithm is a GPU-parallelized version of the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA), using 200,000 dose calculation points (DCPs) and a time limit of 180s (NVIDIA Titan Xp GPU). Resulting plans were re-evaluated on 500,000 DCPs.

To factor in the combined intracavitary and interstitial nature of a cervix BT procedure, interstitial contribution was constrained to be at most 30% of the total dwell time, while single interstitial catheters could contribute up to 20%. The DVI aims of the targets and OARs were taken directly from the EMBRACE-II protocol. Preliminary results showed clearly undesirable locally optimized dose distributions and irregular dwell time patterns. Consequently, the minimum dose to point A was removed from the model, and smoothness in neighbouring dwell times was enforced by a dwell time modulation restriction, where the factor f by which a dwell position with dwell time t can vary from its nearest neighbour is f=2+5/t.

Retrospectively, bi-objective automated treatment planning was done for three patients who had been treated with MRI-guided BT, four fractions of 7Gy HDR each, in two implants. For each of the six BT implants, the dose distributions and DVIs of the treatment plan for which the worst value of the two objectives was best, were evaluated by two medical specialists.

The automatically generated treatment plans were judged as clinically unsatisfactory. The DVIs of the two most criticised plans are given in Table 1. Their dose distributions are shown in Figure 1. The following issues were deemed undesirable: too high doses to normal tissue and outside the CTV_HighRisk, and too low doses to the CTV_IntermediateRisk.

A key advantage of the automated bi-objective approach is being able to quickly generate a diverse set of clinically satisfying plans based on optimization of DVIs. Directly optimizing on the DVIs in the clinical protocol worked well in prostate cancer BT, but, in cervical cancer BT, using EMBRACE-II as the protocol, additional planning aims are required to achieve desirable plans. We are currently developing and adding these to the model, indicating promising improvements.