The Impact of Cardiac Radiation Dosimetry on Survival After Radiation Therapy for Non-Small Cell Lung Cancer

Purpose The heart receives high radiation doses during radiation therapy of advanced-stage lung cancer. We have explored associations between overall survival, cardiac radiation doses, and electrocardiographic (ECG) changes in patients treated in IDEAL-CRT, a trial of isotoxically escalated concurrent chemoradiation delivering tumor doses of 63 to 73 Gy. Methods and Materials Dosimetric and survival data were analyzed for 78 patients. The whole heart, pericardium, AV node, and walls of left and right atria (LA/RA-Wall) and ventricles (LV/RV-Wall) were outlined on radiation therapy planning scans, and differential dose-volume histograms (dDVHs) were calculated. For each structure, dDVHs were approximated using the average dDVH and the 10 highest-ranked structure-specific principal components (PCs). ECGs at baseline and 6 months after radiation therapy were analyzed for 53 patients, dichotomizing patients according to presence or absence of “any ECG change” (conduction or ischemic/pericarditis-like change). All-cause death rate (DR) was analyzed from the start of treatment using Cox regression. Results 38% of patients had ECG changes at 6 months. On univariable analysis, higher scores for LA-Wall-PC6, Heart-PC6, “any ECG change,” and larger planning target volume (PTV) were significantly associated with higher DR (P=.003, .009, .029, and .037, respectively). Heart-PC6 and LA-Wall-PC6 represent larger volumes of whole heart and left atrial wall receiving 63 to 69 Gy. Cardiac doses ≥63 Gy were concentrated in the LA-Wall, and consequently Heart-PC6 was highly correlated with LA-Wall-PC6. “Any ECG change,” LA-Wall-PC6 scores, and PTV size were retained in the multivariable model. Conclusions We found associations between higher DR and conduction or ischemic/pericarditis-like changes on ECG at 6 months, and between higher DR and higher Heart-PC6 or LA-Wall-PC6 scores, which are closely related to heart or left atrial wall volumes receiving 63 to 69 Gy in this small cohort of patients.

carefully chosen. Our work is a first exploratory step and requires further follow up studies to test the hypotheses generated. The cost of these additional studies is low (analysing heart dosimetry and OS data collected in the course of lung RT trials) and the cost of a false negative is high (missing a potentially important discover potentially impacting overall survival), and for these situations a fairly high FDR such as 0.20 is recommended. 1) Bejamini & Hochberg 1995, Journal of the Royal Statistical Society, Series B, Volume 57, No. 1, pp. 289-300. 2) http://www-01.ibm.com/support/docview.wss?uid=swg21476447

Appendix C -Alignment between substructure PCs and heart-PCmax
We define the normalized dot product (NDP) of a substructure PC (ss-PC) as in which j indexes the dose-bins of each PC, and peak-PC is obtained from heart-PCmax by setting all bins outside the high-dose peak of heart-PCmax to zero. An NDP 1 indicates that the high-dose peak component of the substructure PC is at least as great as that of heart-PCmax.
The following table lists the normalised dot products for each of the cardiac substructure PC.

Data preparation
Radiotherapy planning data was exported from the treatment planning system in DICOM-RT format and converted into MATLAB (R2014a; Mathworks, Natick, MA) data-structures using functions from CERR (Computational Environment for Radiotherapy Research) [s1]. Binary label masks, indicating whether each voxel of the CT scan lies inside or outside the structure of interest, were extracted for the heart and cardiac substructures (atria and ventricles) of all patients. Planned dose distributions were interpolated on to the same grid as the CT scan and thresholded at 63 Gy, producing a binary mask of the volume which is dosimetrically similar to the previously-identified dDVH principal component.

Mask registration
One structure set was arbitrarily selected as a reference heart geometry, and a registration procedure was developed to find a coordinate transformation that maps each of the other hearts (the test hearts) on to the reference, as follows. Synthetic registration target images were created from binary label masks for the test and reference hearts, with minor intensity variations introduced using the Euclidean distance transform from the surface of each substructure. An example image is presented in figure S1a). The MATLAB function imregtform was then used to identify the optimal affine transformation (translation, rotation, scale and shear) to map each test image to the reference image. The registration process used mean square difference as an image similarity metric, and a regular step gradient descent optimizer to perform a maximum of 100 iterations at 3 resolution levels. The resulting transformations were also applied to the thresholded dose distribution, allowing cross-comparison of the high dose volume on the geometry of the reference heart. The average final image following affine registration is presented in figure S1b). Figure 2 illustrates the concordance of heart volumes in three dimensions before and after the registration process. The accuracy of the mapping was assessed quantitatively by comparing the masks for the 78 test hearts after transformation to the mask of the reference heart, using a series of standard metrics as given in table S1.

Registration performance
The Dice similarity value indicates that, on average, 90% of voxels are correctly mapped from within the test heart to within the reference heart. The mean distance to agreement between the surfaces of the test and reference hearts is typically less than 3mm, as is the average distance between their centres of mass. Figure S1: a) Synthetic registration target image for the reference heart, created from the delineated heart and substructures. Each substructure (labelled L/R = left/right, A/V = atrium/ventricle) is assigned an intensity, with individual voxel values modulated by the Euclidean distance transform from the substructure surface. The voxel value ranges are distinct for each substructure. Image show an axial slice through the centre of the heart. b) Average synthetic image over all test hearts after registration. Insert: Gaussian kernel (σ = 3mm) used to smooth final dose projections, to same scale as main image. Figure S2: a) Simplified 3D renderings of all hearts in their original coordinate system before registration. Each heart is rendered in a different color. An example heart has been identified by black wireframe rendering. B) Equivalent 3D renderings following the registration process, with the same heart identified by black wireframe. Visual inspection suggests that for this heart the affine transformation constitutes at least translation, scaling and rotation. Table S1: Standard metrics used to assess the performance of the mask registration process The overall performance of the process is considered acceptable, given that affine transformations cannot account for local deformation, although the results indicate that the method is not suitable for identifying features of order 3mm and smaller.

Visualization by projection
Two-dimensional projections were created from each of the registered high-dose volume masks in the coronal, sagittal and transverse planes. The process is equivalent to casting rays in the posterior-anterior, left-right or inferior-superior directions respectively: if a ray encounters one or more voxels of the mask, a '1' is recorded, and otherwise '0' is recorded. Projections in a given plane were summed for all hearts analyzed and the result was divided by the number of hearts, producing a 2D probability distribution for dose 63 Gy occurring at any point on the projection within the structure of interest. The resulting images were convolved with a Gaussian kernel of 3mm standard deviation (shown to scale in the inset of figure 1b) to avoid the identification of features smaller than the typical registration error, and are presented on the left-hand side of each panel in Figure 2 of the main text.
In order to produce the individual substructure images shown on the right-hand side of panels in Figure 2 of the main text, the high-dose mask was first restricted to the appropriate substructure using a Boolean AND operation. The process was otherwise identical.