Exhibit Hall | Forum 2
Purpose: Radiation toxicity of the left anterior descending (LAD) coronary artery is associated with major adverse cardiac events for lung cancer patients treated with radiotherapy. Dose-volume analysis requires to delineate the whole length of LAD, which is challenging in non-contrast CT image, especially for the distal portion of the artery. We propose to use 3D-U Net deep learning with curvature modeling to delineate the whole LAD artery.
Methods: Seventy patients with non-small cell lung cancer (NSCLC) treated with curative-intent radiotherapy were studied. Additional twenty patients with high resolution contrast-enhanced cardiac images were used to model the 3D curvature of the LAD. The proximal, middle, and distal segments of LAD were modeled based on the radius of curvature and direction of curvature vectors. Segmentation of the LAD was constructed by 3D U-Net deep learning. The whole heart contour and planning CT were used as inputs. The cardiac volume was extracted and resized into volumes with 128×128×128 voxels.The curvatures of the LAD segments were derived and integrated with training. Online data augmentation was applied during both training and testing stages. The Dice similarity was used as the loss function. The results were evaluated by mean Hausdorff distance (HD) and Dice similarity coefficients (DSC).
Results: Fifty patients were randomly selected as training dataset, and twenty patients as test dataset. The mean DSC for whole LAD was 0.82+/-0.09, 0.93%+/-0.06 for proximal, 0.82+/-0.08 for middle, and 0.73+/-0.11 for distal segments. The mean HD were 4.3+/-2.1mm for the whole LAD, 1.9+/-0.8mm for proximal, 3.1+/-1.9mm for middle, and 6.5+/-3.1mm for distal segments.
Conclusion: A curvature-guided segmentation was developed to effectively detect and predict the LAD artery in non-contrast CT. The proposed method can be used to identify radiation artery toxicity associated with adverse cardiac events.