Deep learning-based digital subtraction angiography image generation

• Published in: International journal for computer assisted radiology and surgery Vol. 14; no. 10; pp. 1775 - 1784
• Main Authors: Gao, Yufeng, Song, Yu, Yin, Xiangrui, Wu, Weiwen, Zhang, Lu, Chen, Yang, Shi, Wanyin
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.10.2019; Springer Nature B.V; Springer Verlag
• Subjects: Angiography; Angiography, Digital Subtraction - methods; Blood vessels; Computer Imaging; Computer Science; Datasets; Deep Learning; Diagnostic Techniques, Cardiovascular; Digital imaging; Engineering Sciences; Feature extraction; Health Informatics; Humans; Image acquisition; Image processing; Image Processing, Computer-Assisted - methods; Imaging; Machine learning; Medical imaging; Medicine; Medicine & Public Health; Methods; Pattern Recognition and Graphics; Radiology; Review Article; Signal and Image processing; Software reviews; Surgery; Vision; Digital subtraction angiography; Residual dense block; Adversarial network; Artifact elimination
• ISSN: 1861-6410; 1861-6429
• DOI: 10.1007/s11548-019-02040-x

Purpose Digital subtraction angiography (DSA) is a powerful technique for diagnosing cardiovascular disease. In order to avoid image artifacts caused by patient movement during imaging, we take deep learning-based methods to generate DSA image from single live image without the mask image. Methods Conventional clinical DSA datasets are acquired with a standard injection protocol. More than 600 sequences obtained from more than 100 subjects were used for head and leg experiments. Here, the residual dense block (RDB) is adopted to generate DSA image from single live image directly, and RDBs can extract high-level features by dense connected layers. To obtain better vessel details, a supervised generative adversarial network strategy is also used in the training stage. Results The human head and leg experiments show that the deep learning methods can generate DSA image from single live image, and our method can do better than other models. Specifically, the DSA image generating with our method contains less artifact and is suitable for diagnosis. We use metrics including PSNR, SSIM and FSIM, which can reach 23.731, 0.877 and 0.8946 on the head dataset and 26.555, 0.870 and 0.9284 on the leg dataset. Conclusions The experiment results show the model can extract the vessels from the single live image, thus avoiding the image artifacts obtained by subtracting the live image and the mask image. And our method has a better performance than other methods we have tried on this task.