Lung Tumor Motion and its Impact on Deep Learning Prediction of Local Recurrence

Abstract

<h3>Purpose/Objective(s)</h3> In order for deep learning (DL) models to successfully predict local recurrence of lung tumors from image-based features, uncertainties of the model must be quantified. One potential source of uncertainty for the DL model is lung tumor motion. We aimed to evaluate the association of motion magnitude with the prediction accuracy of the DL model. We hypothesized that knowledge of motion will help quantify the uncertainties of prediction and guide the use of individualized risk predictions for lung radiotherapy. <h3>Materials/Methods</h3> Physician delineated gross tumor volume (GTV) of 19 patients were propagated to their corresponding 10 breathing phases of the 4DCT images using a commercial treatment planning system. A programming environment code was used to extract the contours and determine the centroid and volume bounded by the contours. Motion was determined from the displacement of the GTV centroid. The Euclidean distance was calculated and used as an aggregate measure of tumor motion magnitude. The 4DCT, free-breathing CT (FBCT) and maximum-intensity projection (MIP) images of the 18 patients were used as input to a pre-trained DL model. The 8-layer DL model was previously trained to predict local recurrence using FBCT images of 933 lung patients. Correlation between motion magnitude and prediction accuracy was calculated for 4DCT images in each respiratory phase, FBCT and MIP images. <h3>Results</h3> Tumor movement with a mean exertion of 1 cm [IQR: 0.72-1.53] was observed from the cohort with 47.4 % of patients subjected to abdominal compression, and 57.9% of patients had peripheral lung tumors. The standard deviation of DL scores obtained from each 4DCT image correlated with the magnitude of tumor motion (R² =0.45, p=0.005). Higher uncertainty in DL scores was attributed to the larger tumor motion. The average DL score obtained from the 10 phases of the 4DCT images of each patient was (i) highly associated with the DL scores obtained from the FBCT images (r = 0.99; p<0.0001), and (ii) closest to the DL score obtained with the end-inspiration images. In comparing the prediction performance across image acquisition types, MIP images yields a higher DL score compared to FBCT (<i>p</i><0.0001) or 4DCT images (<i>p</i> =<0.0001). <h3>Conclusion</h3> The impact of tumor motion on the range of DL scores for prediction of local recurrence was reported for the first time. Standard deviations of the DL scores increased with motion magnitude. However, when DL score was measured and averaged across the 10 phases of the 4DCT images, DL score correlated well with those obtained from the FBCT images. This preliminary finding suggests that despite the increased variation with larger motion, the DL score, when averaged across the sinusoidal breathing trajectory, correlates with the DL score obtained from the FBCT images that were used for treatment planning. However, caution should be exercised when using MIP images for local recurrence prediction since a higher DL score than the FBCT images used for planning will be observed.