3200 - Geometric and Dosimetric Evaluation of a Hybrid Cardiac Structure Automatic Segmentation Tool for Lung Cancer Radiation Therapy

Purpose/Objective(s):

Prior studies have shown that certain cardiac substructures may be more sensitive to radiation than others. Targeting the radiation dose distribution to avoid the radiation sensitive regions is expected to result in less toxicity. However, accurate delineation of cardiac substructures on planning CT images is a challenging and time-consuming task. Automatic segmentation is needed to consistently contour cardiac substructures more efficiently. Cardiac substructure automatic segmentation tools have been developed and been used in prior studies. The aim of this work was to evaluate a public available automatic segmentation tool for cardiac structures and associated dosimetric accuracies by comparing them with expert contours on a large clinical dataset.

Materials/Methods:

80 local advanced stage lung cancer patients treated with IMRT to a prescribed dose of 45-66 Gy were included in this study. Cardiac substructures, including chambers and coronary arteries were manually contoured by physician experts on planning CT images (mostly Average Intensity Projection of 4DCT). The automated tool utilizes a hybrid technique to combine three automatic segmentation methods. The whole heart is segmented using deep learning neural network U-net, then a multi-atlas image registration guided by heart contour is used to delineate cardiac chambers, and finally coronary arteries are delineated based on geometric modelling. The geometrical accuracies of the automatic segmentations were evaluated by using both volumetric and surface distance metrics, including Dice similarity coefficient (DSC) and mean distance to agreement (MDA). In addition, the mean and maximum dose differences between the manual and automatic substructure contours were calculated using the clinical treatment plans.

Results:

The median (25% to 75% quantile) of DSC for whole heart, chambers, right coronary artery, circumflex artery, LAD are: 0.93 (0.9-0.94), 0.70 (0.63-0.76), 0.05 (0.01-0.09), 0.0 (0.0-0.03), and 0.03 (0.0-0.08), respectively. The values for MDA are (in mm): 2.4 (1.8-3.2), 4.1(3.4-5.5), 8.3 (6.5-11.2), 10.2 (8.7-13.4), and 8.7 (6.9-12.8), respectively. While the dosimetric parameters show reasonable agreements in terms of mean dose differences (in Gy): 0.0( -0.3-0.3), 0.0 (-1.1-0.3), 0.0 (-0.5-1.3), 0.1 (-1.1-3.1) and -0.7 (-3.6-0.1), large deviations were seen in term of maximum dose in the chamber and arteries which has an interquartile range of 7-8 Gy.

Conclusion:

The automatic segmentation tool for heart substructures has reasonable geometric and dosimetric accuracies for whole heart and chambers, and in term of mean dose for other substructures. For coronary arteries large deviations in geometric metrics and maximum dose were observed when evaluated on clinical IMRT cases.