Abstract: FR-PO042
Vascular Calcification Heterogeneity Evaluated by Deep Radiomics Based on Chest Radiography
Session Information
- AI, Digital Health, Data Science - II
November 03, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Author
- Chao, Chia-Ter, National Taiwan University Hospital, Taipei, Taiwan
Background
Vascular calcification (VC) is regarded a systemic pathology involving most arterial segments. Recent studies suggest that calcification heterogeneity exists, meaning that VC involves different wall components and locations, indicating diverse calcification patterns. We previously established a chest radiography-based radiomic approach for identifying VC. In this study, we aimed to evaluate whether radiographic calcification heterogeneity existed, in the form of different VC distributions between those with and without chronic kidney disease (CKD).
Methods
We devised a classification method using the attention mechanism of a deep learning neural network, whose architecture was first trained in a large chest radiography dataset, and then used as an initialization for the target domain with fine-tuning. We divided substrate images into subdivisions with grids and numeric values showing the radiomics features guided by deep learning attention mechanism. Boxes were used to indicate regions more significantly attentioned in patients with at least stage 3b CKD comparing to those in non-CKD populations. We visualized the attention of the network and extracted attentional areas.
Results
We analyzed chest radiography images from 11,106 general population (3.3% with VC) and 59 stage 3b or higher CKD patients (61% with VC). We examined the differentially attentioned areas between CKD and non-CKD patients after deep learning (Figure 1A, overlapping images). Activated areas, or differentially calcified areas in CKD patients included the ascending aorta, carotid vessels, and inter-diaphragmatic areas of thoracoabdominal aortas (Figure 1B). Quantitative analysis revealed inter-diaphragmatic areas exhibited the highest radiomic feature values, followed by carotid vessels and ascending aorta region (Figure 1C).
Conclusion
Using deep radiomics, we demonstrated specific aortic segments and branch arteries to be more significantly affected by VC among CKD patients than those without. Calcification heterogeneity can be detected using deep radiomics.