A weighted feature transfer gan for medical image synthesis
- PDF / 2,144,103 Bytes
- 11 Pages / 595.276 x 790.866 pts Page_size
- 13 Downloads / 186 Views
SPECIAL ISSUE PAPER
A weighted feature transfer gan for medical image synthesis Shuaizhen Yao1 · Jianhua Tan2,3 · Yi Chen1
· Yanhui Gu1
Received: 18 August 2020 / Revised: 10 October 2020 / Accepted: 29 October 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Recent studies have shown that CycleGAN is a highly influential medical image synthesis model. However, the lack of sufficient constraints and the bottleneck layer in auto-encoder network usually lead to blurry image and meaningless features, which may affect medical judgment. In order to synthesize accurate and meaningful medical images, weighted feature transfer GAN (WFT-GAN) is proposed to improve the quality of generated medical image, which is applied to the synthesis of unpaired multi-modal data. WFT-GAN adopts weighted feature transfer (WFT) instead of traditional skip connection to reduce the interference of encoding information on image decoding, while retaining the advantage of skip connection to the information transmission of the generated image. Moreover, the local perceptual adversarial loss combines the VGG feature map and adversarial model to make the local features of the image more meaningful. Experiments in three data sets show that the method in this paper can synthesize higher-quality medical images. Keywords CycleGAN · Medical image synthesis · Weighted feature transfer GAN · Local perceptual adversarial
1 Introduction In recent years, deep learning has been widely used in the medical image analysis and achieved great success [1–4]. Medical image synthesis is one of the hot spots in this field, which has attracted extensive attention of researchers. Medical image synthesis [5–7] plays an important role in clinical diagnosis, treatment planning, and image-guided surgery. The purpose of multimodal medical image synthesis is to synthesize images of one modality into images of another by establishing the spatial mapping relationship between different modalities. Different modal images have different effects
B B
Yi Chen [email protected] Yanhui Gu [email protected] Shuaizhen Yao [email protected] Jianhua Tan [email protected]
1
School of Computer Science and Technology, Nanjing Normal University, Nanjing, Jiangsu Province, China
2
School of Public Administration, Nanjing Agricultural University, Nanjing, Jiangsu Province, China
3
Nanjing Vocational University of Industry Technology, Nanjing, Jiangsu Province, China
on medical diagnosis. However, not all patients will undergo a comprehensive medical examination, which is difficult to obtain enough paired medical data. For example, magnetic resonance (MR) images and computed tomography (CT) [8] images have different advantages in medical diagnosis, and combining the two can effectively assist medical diagnosis. The disadvantage of CT is the lack of good contrast in soft tissues. Radiation exposure during CT collection may also increase the risk of secondary cancer in susceptible people. The difference is that MR can compare soft tissues well. Moreover, c
Data Loading...
