Contextual loss based artifact removal method on CBCT image

Abstract

Purpose: Cone beam computed tomography (CBCT) offers advantages such as high ray utilization rate, the same spatial resolution within and between slices, and high precision. It is one of the most actively studied topics in international computed tomography (CT) research. However, its application is hindered owing to scatter artifacts. This paper proposes a novel scatter artifact removal algorithm that is based on a convolutional neural network (CNN), where contextual loss is employed as the loss function. Methods: In the proposed method, contextual loss is added to a simple CNN network to correct the CBCT artifacts in the pelvic region. The algorithm aims to learn the mapping from CBCT images to planning CT images. The 627 CBCT-CT pairs of 11 patients were used to train the network, and the proposed algorithm was evaluated in terms of the mean absolute error (MAE), average peak signal-to-noise ratio (PSNR) and so on. The proposed method was compared with other methods to illustrate its effectiveness. Results: The proposed method can remove artifacts (including streaking, shadowing, and cupping) in the CBCT image. Furthermore, key details such as the internal contours and texture information of the pelvic region are well preserved. Analysis of the average CT number, average MAE, and average PSNR indicated that the proposed method improved the image quality. The test results obtained with the chest data also indicated that the proposed method could be applied to other anatomies. Conclusions: Although the CBCT-CT image pairs are not completely matched at the pixel level, the method proposed in this paper can effectively correct the artifacts in the CBCT slices and improve the image quality. The average CT number of the regions of interest (including bones, skin) also exhibited a significant improvement. Furthermore, the proposed method can be applied to enhance the performance on such applications as dose estimation and segmentation.

Deep learning methods for enhancing cone‐beam CT image quality toward adaptive radiation therapy: A systematic review

“…The most objective comparisons of image quality can be found in studies that compare multiple architectures or correction techniques for generating sCT images using the same datasets. 18,19,27,35,39,41,44,47,49,57,60,61,63 Where DL methods were compared to classical CBCT correction methods, Barateau et al 61 demonstrated that their GAN sCT achieved a lower MAE than DIR of the CT (82.4 ± 10.6 vs. 95.5 ± 21.2 HU), which was found to be consistent with the results in Thummerer et al (36.3 ± 6.2 vs. 44.3 ± 6.1 HU). 57 Similarly in Liang et al, 39 cycle-GAN showed improved image qual-ity metrics over DIR of the CT when a saline-adjustable phantom was used in a controlled experiment.…”

“…Augmentation of training data is a strategy used to synthetically increase the number of examples to prevent overfitting to the specific variance of the training set and improve model generalizability. The most popular augmentation techniques were random horizontal flips, 18,27,37,42,44,48,55,57,60 followed by random rotations 15,18,44,60,61,63 . The addition of noise, 18,36,46 translational shifts, 55,57,61 random crops, 37,42,44 random shears, 15,61 and scaling 15 were also utilized in the literature.…”

“…The most objective comparisons of image quality can be found in studies that compare multiple architectures or correction techniques for generating sCT images using the same datasets 18,19,27,35,39,41,44,47,49,57,60,61,63 . Where DL methods were compared to classical CBCT correction methods, Barateau et al 61 .…”

“…The most popular augmentation techniques were random horizontal flips, 18,27,37,42,44,48,55,57,60 followed by random rotations. 15,18,44,60,61,63 The addition of noise, 18,36,46 translational shifts, 55,57,61 random crops, 37,42,44 random shears, 15,61 and scaling 15 were also utilized in the literature. Although most augmentation strategies consist of simple linear transformation of the image, the addition of random noise at every iteration degrades image quality yet aids in the learning of salient features by making the network more robust to overfitting (small changes in latent space do not alter the output) while improving generalizability (multiple representations of the same feature are mapped to same output).…”

“…Tables 3a-c show that a total of nineteen studies investigated sCT generation primarily using cycle-GAN for their translation task, 19,[35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51][52] compared to eight utilizing U-Net, [53][54][55][56][57][58][59][60] three implementing GANs, 15,18,61 three exploring deep-CNNs, [62][63][64] and one study utilizing a novel architecture called artifact disentanglement network (ADN). 27 Deep-CNNs maintain the input image dimensions as the feature maps flow through the network.…”