Session Item

Longer overall survival after lung radiotherapy raises concerns about possible treatment-induced cardiotoxic side effects.  Multiple retrospective studies investigated the correlation between heart or cardiac substructure (CS) dose and survival in lung cancer patients. However, due to the general absence of cardiac substructure contours in the clinical care pathway, most studies resort to deforming all dose distributions to a representative anatomy with CS delineations. This study investigates correlation between CS dose and patient survival in a large, single-institute lung cancer cohort with the help of AI-generated CS contours.

We included 1579 lung cancer patients, aged between 36 and 93 years old (median: 70), T-stage 1-4 with median survivals of 51, 39, 21, 18 months respectively, 58% male, that were treated with thoracic radiotherapy between 2009 and 2019. Re-irradiated patients were excluded. A convolutional neural network, trained for this study with 40 patients, created CS contours for each patient. Contouring performance was ensured through random spot checks. As dose effects on organs are traditionally considered as either serial or parallel, D0.1cc (as a surrogate for max. dose) and Dmedian were selected. Planned D0.1cc and Dmedian to CS, the whole heart, and lungs were extracted. Univariant Cox regression was used to find significant correlations between dose and survival. Hereafter, multivariate analysis was used to identify significant correlations, while stratifying for tumor stage, pathology, age, and gender. A Bonferroni-Holm correction was applied.

Univariate Cox regression showed statistical significance for the correlation between dose to all investigated structures and survival except for 3 of the DVH points; left ventricle D0.1cc, Dmedian and left lung Dmedian. On multivariate analysis only DVH points for the left atrium (D0.1cc) and right lung (Dmedian) remained significant (see Table 1). Hazard ratios were calculated for DVH points with significant p-values (p < 0.05) and were 1.003 and 1.009 for left atrium D0.1cc and right lung Dmedian respectively. This correlates any dose to the left atrium with worse overall survival.

Substantial heart dose might not occur in every lung cancer radiotherapy treatment. Verifying the occurrence in this cohort was done by creating a histogram of D0.1cc and Dmedian of the whole heart (See Figure 1). This shows a Dmedian above 10 Gy for 11.8% of the patients, while nearly 2/3rd of the included patients has their D0.1cc above 10 Gy.

Using a set of AI-generated CS contours, and thus avoiding uncertainty associated with deformable dose warping, we showed significant correlation between dose to the right lung and left atrium and survival in a large lung cancer cohort. In future work, we will extend this study to include other cardiac substructures.