Validation of an Image Based Deep Neural Network Model for Individualized Radiotherapy in Lung Malignancies

Abstract

<h3>Purpose/Objective(s)</h3> The current standard of thoracic stereotactic body radiotherapy (SBRT) relies on varied dose prescriptions and is not currently guided by individual tumor features. A deep neural network (DNN) was previously developed that queries thoracic computer tomography (CT)-derived feature space and identifies radiation sensitivity parameters. The DNN can predict treatment failure and provide predictive information for individualization of radiotherapy (RT) dose within a single institution. In this study, the internally tested DNN was applied to an external validation set across distinct systems with the hypothesis that deep learning (DL) scores would continue to predict local failure with a higher C-index than standard radiomic or clinical features. <h3>Materials/Methods</h3> Two distinct cohort-based registries from disparate health systems were used to identify 849 internal and 271 external patients with cancer in the thorax and who were treated with high dose RT using SBRT. Pre-therapy lung CT images from the internal cohort were input to a multi-task DNN to generate an image fingerprint with the primary objective of predicting time to event local treatment outcomes. These findings were subsequently tested in the external cohort. DL scores from each patient were combined with clinical variables to provide an individualized dose for RT that projects a treatment failure probability of <5% at 24 months. <h3>Results</h3> External and internal study populations were significantly different in baseline characteristics, permitting an assessment of model transportability. These included patient demographics, use of motion control, CT scanner model, stage of disease, histology, tumor axial diameter, and dose received (p<0.001). In the external study cohort, radiation treatments in patients with high deep learning scores failed at a significantly higher rate than in those with low scores. The 3-year cumulative incidences of local failure were 25.1% (95% CI: 17.6-33.2) and 10.2% (95% CI: 5.5-16.6), respectively. DL scores independently predicted local failure (hazard ratio 2.16, 95% CI: 1.05-4.44, p = 0.04). A multivariable model that included DL score and cancer types predicted treatment failures with a C-index of 0.71 (95% CI: 0.64-0.76), a significant improvement (<i>p</i> = <0.001) compared to 3D volume and cancer type (C-index: 0.69 [95% CI: 0.63-0.75]). External study patients with recommended RT doses less than or equal to the actual delivered dose had 3-year cumulative incidence of local failure of 0%. Greater discordance between the recommended dose and delivered dose (differences >0 Gy) was associated with higher rates of local failure (<i>p</i> = <0.0001). <h3>Conclusion</h3> The DNN accurately predicted treatment failures from pretreatment CT-images across two distinct clinical settings. Results indicate that our image based DNN can be successfully implemented across other appropriately related populations and provide a robust dose guidance framework for individualized RT in prospective clinical studies.