文献类型： 会议论文

第一作者： Maialen Stephens

作者： Maialen Stephens ¹ ; Raul San Jose Estepar ² ; Jesus Ruiz-Cabello ³ ; Ignacio Arganda-Carreras ⁴ ; Ivan Macia ¹ ; Karen Lopez-Linares ¹ ;

作者机构： 1.Vicomtech Foundation

2.Applied Chest Imaging Laboratory, Brigham and Women's Hospital, Harvard Medical School

3.CIC biomaGUNE, Basque Research and Technology Alliance (BRTA), CIBERES, Ikerbasque, Basque Foundation for Science, Departamento de Quimica en Ciencias Farmaceuticas, Universidad Complutense de Madrid

4.Department of Computer Science and Artificial Intelligence, Ikerbasque, Basque Foundation for Science, University of the Basque Country (UPV/EHU), Donostia International Physics Center (DIPC)

会议名称： International Workshop on Thoracic Image Analysis

主办单位：

页码： 118-129

摘要： Pulmonary Hypertension (PH) is a cardiovascular disease where the pulmonary arterial pressure and resistance increases, causing functional and morphological changes in structures like the Pulmonary Artery (PA). Magnetic Resonance (MR) is a noninvasive image modality whose importance is raising as it allows a complete functional evaluation of the affected structures, as well as complex blood flow pattern analysis. However, anatomical information is more accurately estimated from Computed Tomography (CT), where structures are better defined. Thus, both image modalities are required for a complete PH evaluation and generating CT images from MR seems to be a good alternative to reduce patient radiation, time and expense for a complete characterization of PH only from MR images. Previous approaches have shown CycleGANs to successfully complete image-to-image translation tasks with different medical image modalities, yet, lung MR-CT translation for PA characterization has barely been exploited using CycleGANs. Hence, in this work we propose generating synthetic CT Angiography (CTA) images from pulmonary MR images training a CycleGAN with unpaired data, to further evaluate the ability of the synthetic CT images to provide structural PA information. We demonstrate to generate synthetic CTA images that correctly resemble real CTA images by automatically segmenting the PA with a 2D CNN trained with real CTA scans. We also prove to preserve PA structural information comparing the ground truth PA annotations of the real CTA images and the predicted PA segmentations of the reconstructed CTA scans.

分类号： TP391.41-53