Program Information
A Deep Convolutional Neural Network for Cone-Beam CT Artefacts Reduction
J Xu*, X Han , L Hibbard , V Willcut , Elekta, Inc, Maryland Heights, MO
Presentations
SU-K-FS4-14 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: Four Seasons 4
Purpose: Cone-beam CT (CBCT) plays an important role in traditional and adaptive radiotherapy. CBCT images, however, often suffer from strong scattering and noise artefacts, which prevent their wide-spread use in dose computation, structure delineation, and other image-related tasks. We propose in this work a deep convolutional neural network (DCNN) method to reduce CBCT imaging artefacts to produce high quality CT-like images.
Methods: Inspired by recent developments in deep learning, we propose to improve CBCT image quality by training a DCNN model that can generate CT-like images from given CBCT images. The training data come from existing patients with planning-CTs and daily CBCT images. The training CT and CBCT images are pre-aligned using deformation registration. A fully-mapped DCNN model with 29 layers is then trained to learn a direct mapping from CBCT images to their corresponding CTs. The trained model can be applied on a new CBCT image to generate a corresponding high-quality CT-like image, which effectively removes scattering and noise artefacts from the original CBCT. CBCT and CT image pairs from 21 prostate patients are used as experimental data. CBCT artefacts reduction and quality improvement are evaluated by comparing the automatically generated CT-like images (gCTs) against the real CT images both qualitatively and quantitatively.
Results: By visual inspection, the gCTs showed close resemblance to real CT with imaging artefacts greatly reduced. Quantitative measurements indicated that the gCTs achieved high average structural similarity index (SSIM) values. Intensity profile comparisons also showed that the CT numbers of the gCTs and the corresponding real CT matched closely. Meanwhile, CBCT artefacts reduction through DCNN can be computed very fast using a single GPU.
Conclusion: A novel DCNN method is developed to accurately and efficiently transfer low-quality CBCT images to nearly noise-and-scattering-free CT-like images, which can greatly facilitate the use of CBCT for adaptive radiotherapy.
Contact Email: