Analytical versus Data-Driven Approach of Modelling Brachytherapy Needle Deflection

Abstract

This research is motivated by the need of real-time needle tracking solutions in brachytherapy procedures for improving targeting accuracy. We compared two different modelling approaches to estimate brachytherapy 3D needle deflection during insertion into soft tissue from reaction forces and moments measured at the base of the needle: an analytical model based on beam deflection theory and, a data-driven model using a multilayer perceptron artificial neural network (ANN). Verification of the analytical model as well as training, validation, and testing of the ANN model were performed with experimental data obtained from over 120 insertion tests into gelatine tissue phantoms including a variety of needle types and tissue properties. The ANN model has lower prediction errors and is more robust to changes in testing conditions, with accurate predictions in 3 out of 4 tested scenarios; whereas the analytical model predictions are not statistically comparable to ground truth values in any of the tested scenarios. ANN models show a big potential for online 3D tracking of brachytherapy needles in a clinical context in comparison with beam theory analytical models. A simple neural network trained with numerous needle insertions into representative biological soft tissue could estimate needle tip position with submillimetre accuracy.