Abstract:This study’s purpose is to develop a deep-learning-based approach to generate synthetic MRI (sMRI) for accurate online target delineation and localization and synthetic CT (sCT) for accurate dose calculation during CBCT-guided adaptive prostate cancer radiotherapy (RT). We propose to integrate dense blocks into a 3D cycle-generative adversarial network (cycle-GAN) framework to generate sMRI from daily CBCT, and incorporate residual block minimization into a 3D cycle-GAN framework to generate sCT from daily CBCT. Our networks can automatically learn the transformations from the daily CBCT to the sMRI/sCT through the training CBCT-MRI and CBCT-CT pairs. In addition, the cycle-GANs improve the learning outcomes by incorporating inverse transformations such that the cycle-CBCTs are produced from the sMRI and sCT. To generate synthetic/cycle images similar to the real training datasets, both transformations are implemented by a generator network and their outputs are judged by a discriminator. Training the generator takes into account of the discriminator feedback as well as the estimation errors between the synthetic and real/cycle and real ones. A supervised U-Net has been proposed to segment the prostate from the sMRI generated from daily CBCT. The original MRI-defined prostate was deformed to sMRI assess the prostate segmentation accuracy. Our sCT and CBCT were registered to planning CT (pCT) for generating sCT-based and CBCT-based treatment plans. Mean absolute error (MAE) and normalized cross-correlation (NCC) were used to quantify the differences between the sCT and pCT. Clinically-relevant dose volume histogram (DVH) metrics were extracted from sCT-based and CT-based treatment plans for quantitative dosimetric evaluation. The proposed adaptive RT procedure was tested using 20 prostate cancer patients with 36 treatment plans (mean prescription dose: 45±14 Gy), each of whom has CBCT, MRI and pCT. The mean volume Dice Overlap Coefficient was 89.8±5.5% between our sMRI-based U-net segmented and MRI-defined prostate. The mean MAE and NCC between sCT and pCT were 16.9±5.6 HU and 0.98±0.02. The mean prostate dose differences between sCT- and pCT-based plans were 0.11±0.68%, 0.09±0.41%, 0.04±0.38%, 0.04±0.34%, 0.05±0.38%, and 0.46±1.03% for Dmin, D10, D50, D95, Dmean, and Dmax, respectively. The mean bladder, rectum and femur head dose differences between sCT- and CT-based plans ranged from -0.22% to 0.85% for all DVH metrics. We have developed deep-learning-based methods to generate sMRI and sCT from daily CBCT. This study demonstrated that the proposed methods are capable of reliably improving target delineation and providing comparable dose accuracy to the pCT for prostate cancer treatment planning. The accurate sMRI-based prostate delineation and localization as well as precise sCT-based dose calculation warrants further development of a CBCT-guided adaptive RT workflow for prostate cancer.

A deep learning framework for prostate localization in cone beam CT‐guided radiotherapy

Automated Clinical Target Volume Delineation Using Deep 3D Neural Networks in Radiation Therapy of Non-small Cell Lung Cancer

Automated Clinical Target Volume Delineation for Non-Small Cell Lung Cancer Patients Using Deep 3D Networks

Automatic target positioning and tracking for image-guided radiotherapy without implanted fiducials

Automated contour propagation of the prostate from pCT to CBCT images via deep unsupervised learning

A deep learning-based framework for segmenting invisible clinical target volumes with estimated uncertainties for post-operative prostate cancer radiotherapy

Synthetic CT generation from cone-beam CT using deep-learning for breast adaptive radiotherapy

A Deep Learning Approach to Transform Two Orthogonal X-ray Images to Volumetric Images for Image-Guided Proton Therapy.

Deep Learning-Based Volumetric Image Generation from Projection Imaging for Prostate Radiotherapy.

Cbct-Guided Prostate Adaptive Radiotherapy With Cbct-Based Synthetic Mri And Ct

Cone-beam CT-derived Relative Stopping Power Map Generation Via Deep Learning for Proton Radiotherapy

Deep Learning-Based Relative Stopping Power Mapping Generation with Cone-Beam CT in Proton Radiation Therapy

Daily Cone-Beam CT Multi-Organ Segmentation for Prostate Adaptive Radiotherapy.

Volumetric tumor tracking from a single cone-beam X-ray projection image enabled by deep learning

A Learning-Based Nonrigid MRI-CBCT Image Registration Method for MRI-guided Prostate Cancer Radiotherapy

Deep learning based clinical target volumes contouring for prostate cancer: Easy and efficient application

Deep-learning-based Markerless Tumor Localization Using 2D KV/MV Image.

Deformable MR‐CBCT prostate registration using biomechanically constrained deep learning networks

Deep learning based CT-to-CBCT deformable image registration for autosegmentation in head and neck adaptive radiation therapy

Evaluation of a Deep Learning-Based Pelvic Synthetic CT Generation Technique for MRI-based Prostate Proton Treatment Planning

A deep learning‐based framework (Co‐ReTr) for auto‐segmentation of non‐small cell‐lung cancer in computed tomography images