Graduate School of Informatics, Nagoya University, Chikusa-ku, Nagoya, Aichi, Japan.
Information Strategy Office, Information and Communications, Nagoya University, Chikusa-ku, Nagoya, Aichi, Japan.
Int J Comput Assist Radiol Surg. 2022 Jan;17(1):97-105. doi: 10.1007/s11548-021-02492-0. Epub 2021 Oct 21.
Artery contrasted computed tomography (CT) enables accurate observations of the arteries and surrounding structures, thus being widely used for the diagnosis of diseases such as aneurysm. To avoid the complications caused by contrast agent, this paper proposes an aorta-aware deep learning method to synthesize artery contrasted CT volume form non-contrast CT volume.
By introducing auxiliary multi-resolution segmentation tasks in the generator, we force the proposed network to focus on the regions of aorta and the other vascular structures. Then, the segmentation results produced by the auxiliary tasks were used to extract aorta. The detection of abnormal CT images containing aneurysm was implemented by estimating the maximum axial radius of aorta.
In comparison with the baseline models, the proposed network with auxiliary tasks achieved better performances with higher peak signal-noise ratio value. In aorta regions which are supposed to be the main region of interest in many clinic scenarios, the average improvement can be up to 0.33dB. Using the synthesized artery contrasted CT, the F score of aneurysm detection achieved 0.58 at slice level and 0.85 at case level.
This study tries to address the problem of non-contrast to artery contrasted CT modality translation by employing a deep learning model with aorta awareness. The auxiliary tasks help the proposed model focus on aorta regions and synthesize results with clearer boundaries. Additionally, the synthesized artery contrasted CT shows potential in identifying slices with abdominal aortic aneurysm, and may provide an option for patients with contrast agent allergy.
动脉对比计算机断层扫描（CT）能够准确观察动脉及其周围结构，因此被广泛用于诊断动脉瘤等疾病。为了避免造影剂引起的并发症，本文提出了一种基于主动脉感知的深度学习方法，用于从非对比 CT 容积合成动脉对比 CT 容积。
通过在生成器中引入辅助多分辨率分割任务，迫使所提出的网络专注于主动脉和其他血管结构的区域。然后，辅助任务生成的分割结果用于提取主动脉。通过估计主动脉的最大轴向半径来检测包含动脉瘤的异常 CT 图像。
与基线模型相比，具有辅助任务的所提出的网络具有更高的峰值信噪比值，从而实现了更好的性能。在许多临床场景中应该是主要感兴趣区域的主动脉区域，平均提高可达 0.33dB。使用合成的动脉对比 CT，动脉瘤检测的 F 分数在切片级别达到 0.58，在病例级别达到 0.85。
本研究通过采用具有主动脉感知能力的深度学习模型来解决从非对比到动脉对比 CT 模式转换的问题。辅助任务有助于所提出的模型专注于主动脉区域，并合成具有更清晰边界的结果。此外，合成的动脉对比 CT 在识别腹部主动脉瘤的切片方面具有潜力，可为造影剂过敏的患者提供一种选择。
