A novel deep learning model for enhanced segmentation of internal mammary artery, aorta and their perivascular regions
Sobirov, Ikboljon Xie, Cheng Chan, Kenneth Kotanidis, Christos P. Tomlins, Pete Channon, Keith Neubauer, Stefan Yaqub, Mohammad Antoniades, Charalambos
Sobirov, Ikboljon
Xie, Cheng
Chan, Kenneth
Kotanidis, Christos P.
Tomlins, Pete
Channon, Keith
Neubauer, Stefan
Yaqub, Mohammad
Antoniades, Charalambos
Supervisor
Department
Computer Vision
Embargo End Date
Type
Poster
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Background and Aims: Previous study showed that a new radiotranscriptomic signature, C19RS, has prognostic value in clinical outcomes. Automatic segmentation of vascular structures, such as internal mammary artery (IMA), aorta, and their perivascular regions, from contrast CT angiography (CCTA) would enable high-throughput extraction of the radiomic profile. We aim to develop a novel deep learning model for IMA and aorta that allows automated calculation of C19RS in large cohort analysis.
Methods: The model employs a unique architecture integrating CNN (squeeze-and-excitation block) and transformer (Swin block) to enhance segmentation by alternating the blocks and capturing discriminative features. It was trained on CCTA (n=227) dataset from the OxHVF study in the UK with standardized preprocessing steps (resampling, clipping, and intensity normalization). An iterative model refinement process was applied three times (n=140) to reach a mature model (see figure). An external validation cohort (n=751) from an international site (US) was used, with all segmentations undergoing manual expert review for quality assessment. Finally, a publicly available ASOCA dataset was externally validated (n=318).
Results: The model achieved a mean dice similarity score (DSC) of 0.7876±0.0176 for IMA/peri-IMA segmentation and 0.9207±0.0057 for the aorta/periaortic region segmentation (See figure). The model was refined, reaching 0.947 for the IMA/peri-IMA segmentation. In the external cohort, 679 out of 751 cases (90.4%) were reported clinically acceptable for both regions, with the remaining excluded due to the absence of IMA/aorta in the narrow field of view CCTAs. In the ASOCA cohort, it showed consistent performance (0.961±0.039). These findings highlight the model’s generalizability and scalability for large-scale clinical applications.
Conclusions: This study introduces a robust and clinically adaptable DL model for automated segmenting vascular structures, including IMA, aorta, and perivascular space. Its application in radiotranscriptomic biomarker analysis offers a promising avenue for noninvasive patient outcome predictions, making it a valuable tool for cardiovascular research and clinical practice.
Citation
I. Sobirov et al., “A novel deep learning model for enhanced segmentation of internal mammary artery, aorta and their perivascular regions,” Atherosclerosis, vol. 407, p. 120187, Aug. 2025, doi: 10.1016/J.ATHEROSCLEROSIS.2025.120187
Source
Atherosclerosis
Conference
Keywords
Subjects
Source
Publisher
Elsevier