GANai: Standardizing CT Images using Generative Adversarial Network with Alternative Improvement

Abstract

Computed tomography (CT) is a widely-used diag-1 nostic image modality routinely used for assessing anatomical 2 tissue characteristics. However, non-standardized imaging pro-3 tocols are commonplace, which poses a fundamental challenge 4 in large-scale cross-center CT image analysis. One approach 5to address the problem is to standardize CT images using 6 generative adversarial network models (GAN). GAN learns the 7 data distribution of training images and generate synthesized 8 images under the same distribution. However, existing GAN 9 models are not directly applicable to this task mainly due to the 10 lack of constraints on the mode of data to generate. Furthermore, 11 they treat every image equally, but in real applications, some 12 images are more difficult to standardize than the others. All these 13 may lead to the lack-of-detail problem in CT image synthesis. 14 We present a new GAN model called GANai to mitigate the 15 differences in radiomic features across CT images captured 16 using non-standard imaging protocols. Given source images, 17 GANai composes new images by specifying a high-level goal 18 that the image features of the synthesized images should be 19 similar to those of the standard images. GANai introduces an 20 alternative improvement training strategy to alternatively and 21 steadily improve model performance. The new training strategy 22 enables a series of technical improvements, including phase-23 specific loss functions, phase-specific training data, and the adop-24 tion of ensemble learning, leading to better model performance. 25 The experimental results show that GANai is significantly better 26 than the existing state-of-the-art image synthesis algorithms on 27 CT image standardization. Also, it significantly improves the 28 efficiency and stability of GAN model training. 29 Index Terms-computed tomography, image synthesis, generative 30 adversarial network, alternative training 31

Deep Learning–based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or Masses

“…In our study, applying image conversion was effective, but not a perfect solution in the case of wavelet features. It is possible that further optimized CNN architectures could perform better than our network, such as the image synthetic algorithm proposed by Liang G et al (20). We believe that our study provides a basis for image conversion using CNNs in radiomics and that it will help promote related research.…”

CT Image Harmonization for Enhancing Radiomics Studies

“…In total 14,372 CT image slices from lung cancer patient scans and phantom scans were obtained using three different reconstruction kernels (Bl57, Bl64, and Br40) and four different slice thicknesses (0.5, 1, 1.5, 3mm) using the Siemens CT Somatom Force scanner. We adopted Bl64 kernel as the standard CT imaging protocol, since it has been widely used in clinical practice for lung cancer diagnosis [18]. Two testing datasets were prepared for RadiomicGAN performance evaluation.…”

“…Three state-of-the-art CT image standardization models, i.e. Choe et al [4], GANai [18], and STAN-CT [21], were selected for performance comparison. All the models, including RadiomicGAN, were developed based on TensorFlow [1] and trained using the same training data.…”

“…The model, since it trains a CNN from scratch, requires large training data. Liang et al [18] proposed a cGAN-based [14] CT image standardization model named GANai. A alternative training strategy was developed to effectively learn the data distribution.…”

Cross-Vendor CT Image Data Harmonization Using CVH-CT

“…The model, since it trains a CNN from scratch, requires large training data. Liang et al 8 proposed a cGAN-based 9 CT image standardization model named GANai. An alternative training strategy was developed to effectively learn the data distribution.…”