Introduction
Chronic total occlusion (CTO) is the presence of a complete occlusion in the artery for a minimum of 3 months without any antegrade flow filling indicated by coronary angiography, which is prevalent among patients with ischemic heart disease [
1,
2]. Subtotal occlusion (SO) is severe coronary artery stenosis with the positive flow in the distal segment without complete occlusion [
3]. Distinguishing between CTO and SO before percutaneous coronary intervention (PCI) is clinically relevant because CTO lesions are more difficult to procedure and have a higher rate of late restenosis compared to non-CTO lesions. However, both conditions can present a complete disruption of luminal blood flow on CCTA imaging, making discrimination between the two conditions difficult. Li et al. [
4] and Choi et al. [
3] proposed differentiating CTO and SO based on reverse attenuation gradient (RAG) sign, lesion length, blunt stump and collateral vessels. However, this method largely depends on the type of CCTA scanner used and the radiologist’s experience [
5‐
7]. Thus, searching for simple and easily available indicators for differentiating CTO from SO is required.
Radiomics is a relatively new approach in medicine that uses artificial intelligence-driven analytics to extract and convert digital images into mineable and high-dimensional data for extracting quantitative image features in a high-throughput manner, followed by data analysis to support clinical decision-making [
8]. Radiomics in cardiovascular diseases has recently received much attention, e.g., for identifying features of high-risk plaques, as well as predicting myocardial ischemia and other coronary artery disease [
9‐
12]. A new study shows that a radiomics model can predict the success of percutaneous coronary intervention [
13]. However, there have been no studies on the preoperative application of radiomics to differentiate CTO from non-CTO.
The aim of this study was to develop a diagnostic model to differentiate CTO and SO using non-invasive CCTA imaging-based radiomics.
Discussion
The key findings of this study are: (1) the length and blunt stump were the most sensitive metrics imaging metrics to discriminate CTO and SO; (2) the lumen of the occluded segment of CTO showed different radiomics features compared to SO; (3) radiomics can provide support when the length and shape of the occluded segment are essentially identical.
Differential diagnosis of CTO and SO is clinically important. CTO predicts a more difficult procedure, lower success rate, higher complication rate, higher radiation exposure and longer procedure time for PCI than non-CTO [
2,
28‐
30]. Identification of CTO and SO is probably useful in estimating the difficulty of the procedure or deciding on a revascularization strategy. CCTA is a non-invasive method of assessing coronary artery disease and is recommended as a valuable preoperative imaging tool for CTO [
31].The ability of CCTA to detect CTO may guide more specialized personnel device selection prior to the procedure. Thus, acquiring CCTA image information allows cardiologists to focus on selecting and performing the required procedures without wasting time on diagnosis [
5]. Currently, various CCTA-based clinical evaluation indexes, such as the lesion length, calcification area, presence of blunt stump, and intra-luminal attenuation gradient, are used to differentiate CTO and SO [
3,
4]. However, the acceptable threshold values for the lesion length and TAG are still lacking. Because the non-wide-body detector CT scanner uses a prospective scanning method, differences in the contrast concentration occur in different axial images, which might affect the accuracy of the TAG values [
32]. Additionally, differences in the clinician approach increase the difficulties in assessing TAG and collateral vessels. This also explains why only lesion length and blunt stump resulted to be independent predictors in our study, from which the CCTA imaging model is constructed.
Novel image and data analytic techniques such as radiomics, machine learning (ML), and deep learning(DL) may decrease inter-reader variations, increase the amount of quantitative information, and improve diagnostic and prognostic accuracy while reducing subjectivity and biases [
33]. CCTA provides a platform for linking radiomics to clinical medicine, as it is widely used to diagnose coronary-related diseases because of its low acquisition and post-processing requirements and a large amount of easily available data [
34]. Radiomics can provide information that cannot be perceived quantitatively by human eyes, enhancing our understanding of diseases and ultimately aiding clinical decision-making. The advent of radiomics allows inexperienced clinicians to quickly identify differences that are difficult to distinguish visually. This technique has several benefits, like quick and easy to perform, and requires no additional trauma, exposure to radiation, scanning time, or the contrast agent [
35]. Therefore, our study assists in the manual differentiation of CTO and SO for diagnosis using radiological features, feature selection and construction of predictive models based on machine learning methods.
Recent reports have shown that deep learning models can significantly reduce the post-processing time for CTO quantification on CCTA images compared to traditional manual reconstruction. The occlusion features based on the deep learning model have excellent correlation and consistency compared to the anatomical assessment of manual reconstruction [
25]. Previous radiomics studies on coronary artery disease have focused on the plaque component, peri-coronary adipose tissue or myocardium of the coronary arteries; however, our study first reported using radiomics to a more precise coronary lumen, which may enable the discovery of pathological heterogeneity between CTO and SO. CTO lesions are thrombotic occlusions with fibrous tissues rich in collagen or calcification of the lumen of the occluded segment [
36,
37], whilst SO as an incomplete occlusion. Differentiating CTO and SO based on subjective assessment of CCTA images without considering the complex spatial relationships between voxels may result in the loss of important information. 10 of the 16 features extracted in this study were texture features containing voxels, and the highest coefficient value for model importance was also for texture features. It may potentially reflect differences in pathology, which side-steps the ability of radiomics to discriminate CTO from SO. wavelet_gldm_wavelet-LHH-DependenceVariance and shotnoise_gldm_LargeDependenceHighGrayLevelEmphasis are the two texture features with the highest model coefficient values which respond to the dependence of the gray values in the image. Higher values of the above features indicate that the occluded segment of the vessel has a large dependence of higher gray values, which may be related to the fact that the CTO is enriched with more fibrocalcified tissue [
29,
38].
In this study, we constructed a combined model based on radiomics data and CCTA imaging features of the lumen of the occluded coronary segment. The performance of the combined model in diagnosing CTO from SO was better compared to the imaging and radiomics models, indicating that the former could help overcome the differences in the diagnostic ability of different scanning modalities and inexperienced doctors, resulting in improved diagnostic accuracy. Therefore, the radiomics features based on extracting CCTA images can accurately and reliably distinguish between CTO and SO before PCI and aid clinical decision-making.
The key limitations of this study should be considered. First, this was a retrospective study with small sample size. Second, the perivascular information of the occluded segment was not incorporated because the outlined ROI was located in the lumen, which could result in the loss of some valuable information. Future studies with large sample sizes and prospective designs are needed to validate the model generalizability.
In conclusion, this study develops a diagnostic model to differentiate CTO and SO using non-invasive CCTA imaging-based radiomics that can provide support when the CCTA metrics are similar. Our study would help interventional cardiologists predict the ease of percutaneous coronary intervention. Future studies should assess the value of the radiomics features for guiding treatment.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.