Efficiently train and validate a RapidPlan model through APQM scoring.

PURPOSE

The aim of this study was to propose and validate an intuitive method for training and to validate knowledge-based planning (KBP) systems based on a patient-specific plan quality scoring.

METHODS

A sample of 80 clinical plans of prostate cancer patients were ranked on the basis of the Adjusted Plan Quality Metric (APQM%). This quality metric was computed normalizing the Plan Quality Metric (PQM%) score to the best possible OAR sparing estimated by the Feasibility DVH (FDVH) algorithm. Two different plan libraries were created, purging all the plans below the first quartile or below the median the APQM% distribution. These libraries were used to populate and train two RapidPlan models: respectively, the APMQ and the APMQ models. No further refinements or actions were undertaken on these two models. Their performances were benchmarked against another two RapidPlan models. An Uncleaned model, which was populated and trained with the initial sample of 80 plans, and a Cleaned model, obtained through the standard iterative cleaning and refinement process suggested by the vendor and in literature. The outcomes of a planning test based on 20 patients within the training library (closed loop) and 20 patients outside of the training library (open-loop) were compared through various DVH metrics and the PQM% score.

RESULTS

The selection through APQM% thresholding roughly preserves the geometric variety of the Cleaned model; only the APMQ model showed a modest broadness reduction. The models generated through APQM% thresholding showed target coverage and OARs sparing equal or superior to the Uncleaned and Cleaned models both for the closed- and the open-loop tests. No significant differences were found between the four models. PQM% analysis ranked the overall plan quality as: 86.5 ± 6.5% APQM , 83.1 ± 5.9% APQM , 80.39 ± 10.6% Cleaned and 79.4 ± 8.5% Uncleaned in the closed-loop test; 84.9 ± 7.6% APQM , 82.6 ± 7.9% APQM , 80.39 ± 10.6% Cleaned and 79.4 ± 8.5% Uncleaned in the open-loop test.

CONCLUSIONS

Forward feeding a RapidPlan model through a thresholding selection based on APQM% is proven to produce equal or better results than a model based on a manually and iteratively refined population. A tighter APQM% threshold turns approximately into a higher average quality of plans generated with RapidPlan. A trade-off must be found between the mean quality of the KBP library and its numerosity. The proposed KBP feeding method helps the KBP user, because it makes the model refinement more intuitive and less time consuming.