Room: Stars at Night Ballroom 4
Purpose: Determine if a deep learning based dose predictor (DLDP) for head and neck patients, that can create clinical plans that representing the average quality, can be driven to higher overall plan quality by the use of a quality based self-evolving trainer.
Methods: We simulated the clinical physician decision process by crafting a plan quality score that allowed for direct quantitative comparison of the clinical and DLDP treatment plans. For 126 head and neck patients treated in our clinic, the original clinical plans were compared with the plans predicted using a DLDP. The percent difference in score between the two plans was determined. Patients with a percent difference in the top 80% of the distribution at each iteration were kept in the training set. The validation and testing datasets were not changed through the training process. The new training set was then used to retrain the model for an additional 125 epochs. After retraining, the iterative scoring and training dataset selection was repeated.
Results: Using the testing set as the final comparison between the methods showed that in general the clinical plans scored higher than the baseline predicted plans with a mean score of 30.7 and 19.5 respectively with the mean percentage difference in scores of 42.3%. After the first iteration of retraining the predicted mean score increase from 19.5 to 22.3 with a reduced mean percentage score difference of 30.8%.
Conclusion: We have developed a method to allow a training dataset to self-evolve by iteratively identifying and removing lower quality plans from the dataset. The model trained with the improved dataset can predict dose distributions of higher quality. This self-evolving method allows the dose prediction model to get better over time as higher quality dose plans are created and identified in the clinic.
Not Applicable / None Entered.