PHSOR03 Presentation Time: 9:10 AM

Purpose

Clinical experience or nomograms guide day to day clinical decisions for HDR prostate brachytherapy such as whether there ought to be one more catheter to ensure coverage or whether a given catheter would be unusable as it is too close to urethra. AI or machine learning offers the possibility to mimic this with backing from clinical data. The purpose of this study utilizing simulated data was to explore feasibility of using AI/Machine Learning in answering routine questions during the HDR prostate planning process, such as the number of catheters needed to ensure optimal coverage while ensuring urethral sparing.

Materials and Methods

Data from the catheter insertion and planning during HDR prostate cases such as volume of prostate in the ultrasound and CT, number of catheters are available during insertion and after digitization and optimization the D10% for the Urethra, V150 and V200 for the prostate for a V100 ∼ 95% is obtained. To separate the characteristics of the AI modeling from the peculiarities of the clinical data, a simulated dataset with a gaussian distribution with similar bounds as the typical clinical data was created. AutoML is a subset of machine learning which automates the model validation and evaluation. Using various preset criteria, models are trained on data using fivefold cross validation and a portion of data is held for future testing as a holdout. The scoring metric from this is then used for automatically evaluating the performance of models and choosing the optimal model. Various software solutions were explored for deploying AutoML with low or no code and ability to evaluate the underlying machine learning model predictions being the criterion. The user interface for two of the vendors datarobot and symon.ai were intuitive and easily deployable in comparison to the bigger vendors in the field, of the two, free trial online version of datarobot was used for this study. AutoML was trained and deployed on a set of 51 rows with four of the predictive features - TRUS volume of prostate, number of slices, CT volume and the Dose to 10% of the urethra were used as the training data set for machine learning, with the number of catheters as the target. After completion of the run, the output of top five of the algorithms (elastic net, extreme gradient boosted trees, ridge regressor, light gradient and random forest) were calculated just to evaluate how far off they were from each other and ground truth, using another set of 48 rows of data with some overlap with the training data.

Results

While it was easy to deploy and create a model with this platform, several other platforms such from leaders in the field were much harder to set up and troubleshoot. The outputs for the test data were evaluated relative to the ground truth and the elastic net had the least deviation from the ground truth both in terms of the overall data spread and the deviation from ground truth values. For the cases at the higher end of number of catheters, the predictions deviated significantly more from the ground truth. The mean values of prediction for all the models were close to the mean value of the ground truth and the differences were mostly in the distribution and spread of the data.

Conclusion

AutoML was trained on a small, simulated dataset with 51 rows with four features and tested using another set of 48 rows of features with the target values hidden. The mean predicted values of the catheters were close to the ground truth and the algorithm picked by AutoML did have the least deviation from it. Larger set of data maybe needed to train and eliminate biases. This proof of principle study helped lay out the process for using AI/AutoML on clinical data for this particular question of interest. Future exploration with bigger simulated datasets is underway to understand the systematic biases. Studies with actual clinical datasets are also underway to assess whether the model predictive performance holds clinically.