Radiomic features extracted from magnetic resonance images of a mouse model of sickle cell disease predict genotype and classify arteriopathy using machine learning

Introduction: Sickle cell disease (SCD) is a hereditary blood disorder affecting millions worldwide and is characterized by progressive arteriopathy leading to heightened risk of stroke. Using a label-free magnetic resonance angiography (MRA) method and endpoint histological analysis of the common carotid arteries of a transgenic mouse model of SCD, we determined that SCD causes expansive arterial remodeling in an age-dependent manner [1]. Arterial diameters are only one feature of the SCD-mediated damage visible by human eye, but more information is collected during the MR scans. Radiomics, a growing field in medical imaging, can extract quantitative features from MR scans that, with machine learning, has been used for diagnostic, predictive, and prognostic purposes [2]. Here, we employed a radiomics approach to reveal distinct quantitative features that differentiate between SS mice and their heterozygote (AS) littermate controls. We further hypothesize that these discriminant features will exhibit age-dependent changes, reflecting the worsening vascular phenotype of SCD with age or the subject-to-subject variability in disease severity that also complicates clinical treatment of the human condition. Our aim is to identify key radiomic features that could identify those mice that will later develop the more severe phenotypes of SCD, or even AS mice that will experience greater complications than the perceived benign sickle cell trait using this non-invasive diagnostic approach.

Materials and

Methods: We used a humanized Townes sickle cell transgenic mouse model that is homozygous for β-globin S mutation (SS) and heterozygous (AS) as controls. Male and female mice were examined at 1 and 3 months of age. All experiments were approved by Georgia Tech’s IACUC. Mice were phenotyped for sickle (HbS) and normal human hemoglobin (HbA) by native hemoglobin polyacrylamide gel electrophoresis of their blood obtained via paw needle puncture. A total of 66 MRI scans of carotid arteries were collected from one-month-old mice (n=28 AS; n=38 SS) and 64 MRI scans from three-month-old mice (n=33 AS; n=31 SS). Image processing was performed via Mimics Materialise software and involved manual segmentation and 3D reconstruction of the common carotid arteries for region of interest (ROI) delineation and morphological assessment of the arteries (Fig. 1A). All vessels were removed except the common carotid arteries which were then cleaned as they branch from the ascending aorta to the carotid bifurcation. Centerline measurements of the cross-sectional area along the length of the whole artery were made at a resolution of 0.15 mm. 112 Radiomic features were extracted from each single ROI using PyRadiomics software. This was followed by the selection of the most discriminating features, training of a predictive model based on discriminant analysis, and then predictive model validation to differentiate between SS and AS mice at both time points.

Results, Conclusions, and Discussions: Radiomics analysis revealed discriminant features for differentiating carotid arteries in SS and AS mice. At one month, a combination of four features yielded an area under the receiver operating characteristic curve (AUROC) of 0.72; 95% C.I. between 0.60 and 0.84 with p value of 0.032, and an accuracy of 74% (43/66 scans were correctly identified) (Fig. 1B). These features included original gray-level co-occurrence matrix (GLCM) IMC1, first-order minimum, minor axis length, and major axis length. At three months of age, only one first order texture feature, mesh volume, was significantly different between the AS and SS mice. This single feature achieved accurate diagnostic performance in discrimination between SS and AS, with an AUROC of 0.67; 95% C.I. between 0.61 and 0.87 with p value of 0.002 and an accuracy of 76% (48/64 scans were correctly identified) (Fig. 1C). These findings suggest the potential of radiomics in identifying MRA quantitative features that indicate SCD-mediated arterial remodeling and progression with age. There were some outlier cases that were incorrectly identified based on radiomics predictions that could indicate discrepancies between the predicted features and the actual biological conditions. For instance, upon examining MRA and 3D reconstruction images, as well as the cross-sectional area measurements of the incorrectly predicted mice, abnormalities such as stenosis and arterial tortuosity were evident in some MRA scans, contributing to the discrepancy in quantitative predictions (Fig. 2).

This is the first study to demonstrate the feasibility of radiomics analysis in discriminating common carotid artery MRA features between SCD and control mice and predicting their genotype. The identified distinct radiomic features might be associated with arterial physiological changes associated with SCD at different disease stages. MRA imaging along with the use of radiomics and advanced machine learning techniques, offer a non-invasive approach to assess vascular changes in preclinical models of SCD. Further research is warranted to validate these findings in clinical settings and explore the prognostic utility of translating radiomics in SCD management in humans.

Acknowledgements (Optional):

References:
[1] Song, H., Keegan, P. M., ... & Platt, M. O. (2020). ATVB, 40(5), 1220-1230;
[2] Comelli A., Stefano A., … & Barone S. (2020). Medical Image Understanding and Analysis Conference (MIUA)