Automating the aortic valve calcium scoring process
Challenge
Aortic stenosis stands out as the predominant primary valvular ailment necessitating surgical or transcatheter intervention in developed countries. The primary test used to diagnose this pathology is Cardiac computed tomography (CCT). CCT is employed to quantify aortic valve calcification, serving to evaluate the severity of aortic stenosis, monitor disease advancement, and predict major cardiovascular events.
However, calcium deposits often manifest in various areas of the aorta and heart, resulting in misleading regions and an inaccurately computed aortic valve calcium score. In one of our projects, we addressed the challenge of eliminating false-positive regions (calcifications detected outside the aortic valve) from CCT scans. We implemented a Particle Swarm Optimization (PSO) algorithm specifically designed for this purpose.
What we did
From a clinical point of view, quantification of aortic valve calcification is proving valuable for assessing the degree of aortic valve stenosis in cases where Doppler echocardiography, the primary method of assessment, does not yield conclusive results.
Manual segmentation of aortic calcifications limits medical exam speed and accuracy, which has become an important challenge due to the growing importance of aortic valve calcium scoring in clinical practice.
To identify potential calcifications of the aortic valve, our project utilizes anatomical information about the aorta using a deep neural network for aorta segmentation (Fig. 1). This network is then applied to a contrast CT scan co-registered to the corresponding non-contrast image. Thus, we obtained high-quality segmentation of the aortic root and aortic valve. The next step is to implement a PSO algorithm to remove calcifications that are located outside of the valve and may impact the aortic valve calcium score.
Fig. 1: A flowchart of the automated aortic valve calcium scoring.
We conducted a study on 30 pairs of contrast and non-contrast CCT scans to demonstrate the effectiveness of our end-to-end approach. We obtained the images from three different institutions that used scanners from four different manufacturers. Our approach was validated by a medical expert with 18 years of experience who assessed the quality of segmentation. The scans come with segmentation masks for the aorta and aortic root.
The result
Precise calculation of aortic valve calcification from CCT in a reproducible, user-independent, and unbiased manner holds clinical significance, facilitating improved patient selection for surgical interventions. Addressing this challenge, we proposed a comprehensive end-to-end procedure for task automation. This approach incorporates a cascaded deep learning method for aorta segmentation, followed by thresholding the aorta’s 3D mask to identify calcification candidates. Further, we used a Particle Swarm Optimization (PSO) technique to eliminate calcifications positioned outside the valve (Fig. 2), ensuring an accurate aortic valve calcium score. The experimental study showed that the results obtained for all of the analyzed 30 CCT scans were assessed as good or very good by an experienced clinical expert, showing its potential practical utility.
Fig. 2: Results from (a) PSO, rated as either good (G) or very good (VG) by a reader, and the corresponding (b) GSF (with calcifications rendered in yellow).
A detailed description of the project can be found in the following paper: Jaroslaw Goslinski, Filip Malawski, Mariusz Bujny, Marcin Kostur, Karol Miszalski-Jamka, Jakub Nalepa, Deep learning meets particle swarm optimization for aortic valve calcium scoring from cardiac computed tomography, 2023 IEEE International Conference on Image Processing (ICIP), DOI: 10.1109/ICIP49359.2023.10223100.