H. Bacon1, N. McNeil2, T. Patel3, M. Welch4, X. Y. Ye5, A. Bezjak1,6, B. H. Lok1,4, S. Raman1,6, M. E. Giuliani1,6, J. Cho7, A. Sun1,4, G. Liu8, S. Kandel9, C. McIntosh1,10, T. Tadic1,4, A. J. Hope4,7, and P. E. Lindsay Jr1,4; 1Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada, 2Calvary Central Districts Hospital, Elizabeth Vale, Australia, 3Techna Institute, University Health Network, Toronto, ON, Canada, 4Radiation Medicine Program, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, 5Department of Biostatistics, Princess Margaret Cancer Centre, Toronto, ON, Canada, 6Radiation Medicine Program, Princess Margaret Cancer Centre, Toronto, ON, Canada, 7Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Toronto, ON, Canada, 8Division of Medical Oncology and Hematology, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, 9Department of Medical Imaging, University of Toronto, Toronto, ON, Canada, 10Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada
Purpose/Objective(s): Radiation pneumonitis (RP) is a common and potentially severe complication following radiotherapy for locally advanced non-small cell lung cancer (NSCLC). We previously reported on an artificial intelligence (AI) algorithm to identify patients with radiographic features of interstitial lung disease (ILD), and demonstrated that these patients are at greater risk of developing severe grade =3 (G3+) RP. In this work we evaluate dosimetric predictors of G2+ and G3+ RP in patients identified by our AI-ILD algorithm. Materials/
Methods: All locally advanced NSCLC patients treated with conventionally-fractionated definitive radiotherapy at our institution from 2006-2021 with full dosimetric information available were assessed. RP data were prospectively collected and retrospectively reviewed. Dose volume histogram (DVH) data and clinical data were extracted from an institutional database. A convolutional neural network was previously developed and validated to identify patients with radiographic ILD. Planning CT scans for the retrospective cohort were used as input to the algorithm, with AI-ILD score as an output. AI-ILD scores above our established threshold were labeled as AI-ILD+. Univariate correlations between RP and dosimetric parameters including MLD (mean lung dose), Dx (minimum dose to x% volume) and Vx (% volume receiving at least xGy dose) were calculated using Spearman coefficients (R). Results: 498 patients were available for analysis. Of these, 141 (28%) were identified as AI-ILD+. Grade 2-5 RP was reported in 17%, 4.8%, 0% and 0.60% of all patients respectively. Grade 2-5 RP was reported in 14%, 7.1%, 0%, and 1.4% of AI-ILD+ patients respectively, compared to 18%, 3.9%, 0% and 0.28% of AI-ILD– patients. There was a higher MLD in the AI-ILD+ cohort (mean 16Gy, median 16Gy, IQR 13Gy-19Gy) compared to the AI-ILD– cohort (mean 15Gy, median 15Gy, IQR 12Gy-18Gy). Our previous multivariable analysis incorporating MLD showed a significant correlation between G3+ RP and AI-ILD status. When cohorts were dichotomized by AI-ILD threshold, we identified higher correlations between dosimetric parameters and G2+ and G3+ RP in AI-ILD+ patients. The Spearman coefficient (R) for the correlation between MLD and G2+ RP in AI-ILD+ patients was R=0.25 (p=0.0015), compared to R=0.092 (p=0.041) in AI-ILD– patients. Similarly for V20 and G2+ RP, R=0.25 (p=0.0016) for AI-ILD+ patients, compared to R=0.092 (p=0.040) for AI-ILD– patients. AI-ILD– patients demonstrated no significant correlations between MLD or V20 and G3+ RP. For AI-ILD+ patients, both MLD (R=0.17, p=0.020) and V20 (R=0.14, p=0.044) were significantly correlated with G3+ RP. Conclusion: In summary, our AI-ILD algorithm identifies a subset of patients who have a greater correlation between dosimetric parameters and the development of G2+ and G3+ RP. This may motivate the application of lower dose constraints in this population to mitigate the increased risk of severe RP.
