Deep Learning-Based Pseudo CT Reconstruction for MR Only-Guided Radiation Therapy of Lung SBRT

Abstract

The improved soft tissue contrast offered by magnetic resonance imaging (MRI) systems compared to traditional computed tomography (CT) systems makes MR only-guided radiation therapy (RT) an attractive goal for future RT applications. However, a major limitation in moving towards this goal is the necessity of the electron density information derived from CT in RT treatment planning for dose calculations. To overcome this problem, we propose a deep learning-based pseudo CT reconstruction method using a generative adversarial network (GAN) and investigate its use in MR-guided RT treatment planning for lung SBRT. In this study, we have used a GAN to reconstruct pseudo CT images for treatment planning in MR-guided RT of lung SBRT. A GAN consists of two competing networks: 1) a generator that produces pseudo CT outputs based on MR image inputs and 2) a discriminator that differentiates between real and generated images. The networks are trained iteratively to improve the capabilities of the discriminator and the quality of the generator outputs. The generator uses a U-net architecture of 16 convolutional layers with skip connections that transfer high resolution features. The discriminator consists of 5 convolutional layers. Registered MR and CT images from 37 patients previously treated at our institution were used to train the network. The images were masked to eliminate background elements outside of the body contour. Also, the ranges of pixel values in both image sets were restricted to create a more homogeneous training set. After training, the model was applied to unseen MR images from 4 patients for testing. Electron density information derived from the reconstructed pseudo CTs was used to compute the dose based on the same plan parameters used in the clinical plan. Dose-volume histograms (DVHs) were used as the basis of comparison between the proposed and clinical plans. A total of 2000 training epochs took approximately 6 days to complete. When applied to new image sets, the trained model had a throughput time of approximately 0.1 s/slice. DVHs based on the pseudo CTs were comparable to those of the clinically delivered plans. Calculated dose metrics for the proposed and clinical plans are summarized in Table 1 along with 3D Gamma Index pass rates. The comparable calculated dose demonstrates the efficacy of pseudo CT images in treatment planning. Adopting the proposed method offers the potential to eliminate extra radiation dose from unnecessary CT scans in a move towards MR-only RT.Abstract MO_28_2725; Table 1PTVLungEsophagusHeartSpinal Cord3D Gamma IndexPt.D95 (%)V20 (%)Dmax (Gy)Pass Rate (%)1Clinical99.9116.3169.6335.6921.5599.87Proposed99.9716.4169.7536.4221.842c.99.236.5032.8761.4826.9798.14p.99.076.5332.5160.7027.173c.68.5016.1743.3165.2116.9399.86p.67.5116.1442.7765.0416.974c.99.999.1063.1763.3725.1098.33p.100.008.8062.6862.6825.94 Open table in a new tab