Towards Automatic Identification of Missing Tissues using a Geometric-Learning Correspondence Model

TL;DR

The paper tackles the challenge of identifying missing tissue in intra-patient head-and-neck meshes to improve dose mapping in reirradiation. It introduces a pipeline that leverages a pre-trained geometric-learning correspondence model and uses the forward-backward prediction discrepancy, quantified as the correspondence-based inverse consistency error ($cICE$), to locate missing tissue; the threshold $t$ is optimized to maximize predictive accuracy. Using 35 simulated mandible resections, the authors achieve a balanced accuracy score of $0.883$ with an ensemble approach at $t=5.5$ mm, and demonstrate plausible results in a real case with ~25% mandible loss, though the method fails for an extreme ~50% loss. The work highlights the potential of geometry-driven contour registrations for radiotherapy dose estimation in reirradiation, while acknowledging limitations in extreme cases and the need for adaptive strategies and additional imaging data for robustness.

Abstract

Missing tissue presents a big challenge for dose mapping, e.g., in the reirradiation setting. We propose a pipeline to identify missing tissue on intra-patient structure meshes using a previously trained geometric-learning correspondence model. For our application, we relied on the prediction discrepancies between forward and backward correspondences of the input meshes, quantified using a correspondence-based Inverse Consistency Error (cICE). We optimised the threshold applied to cICE to identify missing points in a dataset of 35 simulated mandible resections. Our identified threshold, 5.5 mm, produced a balanced accuracy score of 0.883 in the training data, using an ensemble approach. This pipeline produced plausible results for a real case where ~25% of the mandible was removed after a surgical intervention. The pipeline, however, failed on a more extreme case where ~50% of the mandible was removed. This is the first time geometric-learning modelling is proposed to identify missing points in corresponding anatomy.

Figures (6)

• Figure 1: Example of input meshes and the derived $cICE$. Notice that large $cICE$ values are located in regions of point mismatch between the source and target meshes.
• Figure 2: Simulated cuts for the CT2 mandible of a patient.
• Figure 3: Golden standard for the mesh pair shown in figure \ref{['fig:cICE']}(a), where the dark points have been identified as missing. (b-d) points identified as missing after applying the proposed pipeline for different thresholds and their balanced accuracy score (BAS). These thresholds are applied to the $cICE$ presented in figure \ref{['fig:cICE']}(b).
• Figure 4: Balanced accuracy score (BAS) for all simulations and tested thresholds, discriminating by single model vs ensemble approach (top) and by patient (bottom).
• Figure 5: Performance of the ensemble pipeline using the found optimal threshold of 5.5 mm to data of two patients who underwent mandible surgery due to osteoradionecrosis after their first radiotherapy course.