Hybrid Deep Learning and Model-Based Needle Shape Prediction

Needle insertion using flexible bevel tip needles are a common minimally-invasive surgical technique for prostate cancer interventions. Flexible, asymmetric bevel tip needles enable physicians for complex needle steering techniques to avoid sensitive anatomical structures during needle insertion. For accurate placement of the needle, predicting the trajectory of these needles intra-operatively would greatly reduce the need for frequently needle reinsertions thus improving patient comfort and positive outcomes. However, predicting the trajectory of the needle during insertion is a complex task that has yet to be solved due to random needle-tissue interactions. In this paper, we present and validate for the first time a hybrid deep learning and model-based approach to handle the intra-operative needle shape prediction problem through, leveraging a validated Lie-group theoretic model for needle shape representation. Furthermore, we present a novel self-supervised learning and method in conjunction with the Lie-group shape model for training these networks in the absence of data, enabling further refinement of these networks with transfer learning. Needle shape prediction was performed in single-layer and double-layer homogeneous phantom tissue for C- and S-shape needle insertions. Our method demonstrates an average root-mean-square prediction error of 1.03 mm over a dataset containing approximately 3,000 prediction samples with maximum prediction steps of 110 mm.