Computationally Efficient Method for Estimating the Impact of Cardiorespiratory Motion on Patients Undergoing Stereotactic Arrhythmia Radioablation

Abstract

As stereotactic arrhythmia radioablation (STAR) continues to be studied and additional clinical outcomes become available, the importance of accurate measurements or clinically acceptable estimates of delivered dose, accounting for both cardiac and respiratory motion, becomes paramount during treatment planning to inform clinical decisions. The impact of patient-specific cardiac and respiratory motion on radiation dose delivery has been reported in a limited number of studies. Still, the workflow required to recalculate this dose is highly time-consuming and clinically infeasible during treatment planning. This study aims to (1) measure the impact of cardiac and respiratory motion on the patients treated in our cohort, as well as (2) find computationally efficient estimations of the delivered dose to allow incorporation of cardiac and respiratory motion into clinical treatment planning decisions. To accomplish aim 1, a set of CT images covering all cardiorespiratory phases needs to be created from the initial planning CT to recalculate the delivered dose. To achieve this, we measured the cardiac and respiratory motions separately from the cardiac and respiratory 4DCT images. Metal artifacts, from implantable cardioverter-defibrillator (ICD) lead wires, hindered direct registration of the 4DCT images. To address these artifacts, a U-Net-based deep-learning model for artifact segmentation was developed to identify the ICD leads automatically. A second deep learning network was created using a generative adversarial network (GAN) to use the artifact segmentations and surrounding image information to replace the artifact-affected region with anatomically consistent values, enabling more accurate registration of the 4DCT images. The resulting DVFs from the registration of the cardiac and respiratory 4DCT images were applied to the initial planning CT image to create 100 cardiorespiratory phase images. The initial treatment plan was recalculated on each of these images and deformed using the inverse DVFs to align the doses anatomically to be averaged. To accomplish aim 2, we use the 100-phase dose calculated in aim 1 as a reference and explore the results of reducing the number of phases used for calculations. Three potential avenues were explored: 1) decreasing the number of cardiac phases considered from 10 to 3 while keeping all 10 respiratory phases, 2) decreasing the number of cardiac and respiratory phases from 10 to 3, and 3) using only the initial dose map but adding the dose from the surrounding area proportional to the probability of the voxel being in that location which we called spatial probability map (SPM) dosimetry. The best agreement to the 100-phase dose came from the SPM dose, with the 30-phase dose also having clinically acceptable agreement. The 9-phase dose matched the 100-phase dose better than the initial 3D calculation but did not meet our acceptable replacement criteria. We believe that the SPM dosimetry methodology can be useful in predicting the impact of cardiorespiratory motion and utilized in a clinical workflow to analyze proposed plans for their target coverage without requiring additional manual recalculations. Retrospectively analyzing eight patients with the SPM dosimetry, and the 30-phase dosimetry as a backup, we found that five patients no longer met clinical target coverage goals.