Extracting radiotherapy treatment details using neural network-based natural language processing

Abstract:

Results Performance for each model is below. The best result was for RT dose, likely because of its standard, unique reporting. The worst results were for boost, likely due to few entity instances. The F1 score approximated or exceeded the inter-annotator agreement, which serves as the system target, for RT dose, fraction number, and treatment site. Conclusion Taking advantage of language models pre-trained on non-medical text and using contextual word and character embeddings, neural networks achieved reasonable performance on RT detail extraction despite the small dataset and the highly specialized language. Ongoing efforts include expanding the dataset, and developing models for additional RT entities and for relations between RT details.