Segmentation of Targets and Organs at Risk for CBCT-Based Online Adaptive Radiotherapy Using Recurrent Neural Networks: A Clinical Evaluation

Abstract

<h3>Purpose/Objective(s)</h3> Accurate and automated segmentations of targets and organs at risk (OARs) are critical for the successful clinical implementation of online adaptive radiotherapy (oART). Deforming contours from an initial planning CT to a newly acquired CBCT is one popular approach utilized by major oART platforms to facilitate online delineation. However, this can be challenging due to poor quality CBCT imaging or large anatomical changes from simulation to treatment. In this work, we developed a framework to utilize contours from previous adaptive treatments (per patient), to further improve the accuracy of the auto segmentations for the upcoming fractions (FXs). <h3>Materials/Methods</h3> Based on a 3D U-Net, we built a Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM). The most recent prior FX contours and the base U-Net were used as performance benchmarks. The model inputs included an image of the most recent FX, its clinician-edited contours, as well as an image of the current FX. The label was the clinician-edited contours for the current FX. For each patient, we trained the model one FX at a time via sequential FXs and reset the memory after the last FX. The memory of LSTMs within each down-sampled block preserves a multi-scale memory about the current patient. The models were pre-trained using simulated data. The simulated data was generated from 95 Head & Neck patients at our institution, by interpolating and extrapolating the deformation vector field from planning CT to CBCT image. We fine-tuned and tested our models using a Head & Neck dataset from a clinical oART system. Contour accuracy was measured using DICE. <h3>Results</h3> The clinical oART dataset included 10 patients and each patient had 6 FXs. Four patients were used for fine-tuning, 2 for evaluation, and 4 for testing. Structures of 4 OARs and GTV-nodal are tested. RNN with prior FX knowledge improved the quality of contours as shown by the average DICE scores in Table 1. "Input" column represented using the prior FX edited contours as the current FX contours. CNN represents the base U-Net model. <h3>Conclusion</h3> Automatic segmentation of OARs and targets on CBCT images can be challenging. By utilizing the contours from previous FXs as prior knowledge through an RNN architecture, we can achieve higher accuracy for segmenting OARs and targets in CBCT images for online adaptive radiotherapy.