A Leaning-Based Method to Improve Cone Beam CT Image Quality for Adaptive Radiation Therapy

Abstract

Quantitative Cone Bean CT (CBCT) imaging is on increasing demand for precise image-guided radiation therapy (RT) because it provides a foundation for advanced image-guided techniques, including accurate treatment setup, online tumor delineation, and patient dose calculation. With more precise treatment monitoring from CBCT, dose delivery errors can be significantly reduced in each fraction or compensated for in subsequent fractions using adaptive RT. However, the current CBCT has severe artifacts mainly due to scatter contamination and its current clinical application is therefore limited to patient setup based only on bony structures. This study’s purpose is to develop a learning-based approach to improve CBCT image quality for quantitative analysis during adaptive RT. The first step is to build a set of paired training images including planning CT and CBCT. For each pair, the planning CT is used as the regression target of the CBCT. We then remove the uninformative regions, reduce noise, and perform an alignment between CT and CBCT. The proposed correction algorithm consists of two major stages: the training stage and the prediction stage. During the training stage, patch-based anatomical features are extracted from the registered training images with patient-specific information, and the most robust and informative CT-CBCT features are identified by feature selection to train random forests. During the correction stage, we extract the selected features from the new (target) CBCT and feed them into the well-trained forests to predict the corrected CBCT. This prediction-based correction algorithm was tested with brain CBCT and CT images of 9 patients. We performed leave-one-out cross-validation method to evaluate the proposed prediction-based correction algorithm. The mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and feature similarity (FSIM) indexes were used to quantify the differences between the corrected CBCT and planning CT. For 9 patients, the mean MAE, PSNR, and FSIM were 5.87±4.39, 28.47±4.71, and 0.90±0.08, respectively, which demonstrated the corrected CBCT prediction accuracy of the proposed learning-based method. To improve CBCT imaging, we have developed a novel learning-based method in which a random forest regression with a patch-based anatomical signature is used to effectively capture the relationship between the planning CT and CBCT. We have demonstrated that this method could significantly reduce scatter artifacts. The proposed method has great potential in improving CBCT image quality to a level close to planning CT, therefore allowing its quantitative use in adaptive RT.