Room: AAPM ePoster Library
Purpose: To develop a machine learning (ML) algorithm for predicting the use of hybrid interstitial (HIS) needles for high-dose-rate (HDR) cervical brachytherapy based on high-risk clinical target volume (HR-CTV) spread metrics.
Methods: A dataset of 86 fractions using HIS applicators was employed to train and validate a supervised multi-label classification ML algorithm. The mean and maximum HR-CTV lateral and vertical extent, volume, and offset of the HR-CTV from the applicator tandem axis were extracted from patient contours. These features were divided into 33° sections centered over the needle channels (labelled 4-12) of the Vienna applicator (Elekta AB, Stockholm, Sweden) to express the directional HR-CTV geometry. A k-nearest neighbors algorithm was selected to predict the use of HIS needles based on these directional features. The process of algorithm training/testing was repeated for 1,000 random selections of training/testing data (75%/25%) to evaluate performance. This dataset suffered from label imbalance as certain needle positions were used less frequently. To account for this imbalance, model performance was evaluated using micro-averaged metrics that calculate the metric globally.
Results: The F1 score – the harmonic mean of precision and recall - was 83.9%±3.5%. The Hamming loss - a measure of the disagreement between the needles predicted by ML and the clinical selection - was 12.8%± 2.8%. Individual needle performance was calculated with F1 scores for needles 4(31.8%), 5(90.1%), 6(89.7%), 7(64.3%), 8(0.0%), 9(64.1%), 10(87.0%), 11(89.7%), and 12(40.0%). Lower F1 scores for certain needles can be attributed to the infrequent use of these needles clinically.
Conclusion: This work presents an early proof of concept for a needle selection algorithm. This algorithm demonstrated high predictive accuracy, with a majority of incorrect predictions coming from infrequently used needles. This illustrates the potential for ML to be a powerful predictive tool for needle selection, but highlights the need for more data.