Flow Profile Characterization Using 2D PC MRI Radiomics in Aortic Stenosis

Background: Sex-specific differences in patients with aortic stenosis (AS) are subject to ongoing research. While distinct pathogenesis, presentation and prognosis have been reported¹, and 4D phase-contrast (PC) MRI patterns described², a characterization in 2D PC MRI, a faster technique to quantify aortic flow, is desirable. We aim to differentiate flow patterns in AS patients using radiomics extracted from 2D PC MRI and study interactions with clinical profile and proteomic expression.

Methods: We analyzed 2D PC MRI velocity maps of AS patients before aortic valve replacement surgery³, and of healthy controls⁴.

The ascending aorta (AAo) in a plane just distal of the ostia was segmented in corresponding magnitude images. To minimize effects of imaging resolution, each map was resampled to an individual voxel size to yield 605 voxels within the AAo. We extracted 101 radiomics features from the AAo region at peak systole, and 10 features quantifying throughflow, flow eccentricity and relative aortic movement.

Distinct flow patterns were discovered with non-negative matrix factorization (NMF) on normalized radiomics features of the AS cohort and the resulting groups compared. In addition to clinical parameters, we analyzed proteomic profiles of an AS subcohort, obtained from left-ventricular biopsies during valve replacement, and non-AS controls as previously described³.

Results: We included 60 AS patients and 19 age-matched healthy controls in the 2D PC MRI analysis (Tab. 1). 41 of AS patients had proteomic profiling, along with 17 non-AS controls (distinct from imaging controls).

Resampling to uniform AAo voxel count decouples radiomics parameters that are resolution-sensitive from patients’ aortic size (Fig. 1b). While men and women did not differ in AAo area in our study, size-sensitive resampling precludes confounding radiomics analyses by the choice of FOV and voxel-size relative to the size of the aorta.

NMF resulted in two stable clusters. Flow profiles of patients in cluster A show more homogeneous flow velocities, but higher eccentricity, commensurate with mostly bicuspid valves. Cluster B is characterized by central peak flow, higher backflow and heterogeneous velocity distributions (Fig. 1c).

Clusters do not align with sex: cluster A contains more women, cluster B more men. Men in A have low BSA and high AAo area, women in B vice versa. Patients in B have lower ejection fractions, higher heart failure markers, and more ventricular scar tissue.

Of proteins relevant in AS³, serpinE2, involved in cardiac fibrosis, is upregulated in men and in cluster B women, while STAT3, relevant in cardiac remodelling, is differentially downregulated between A and B in both sexes.

Conclusion: Radiomic profiling of 2D PC MRI velocity maps identifies two AS subgroups. Notably, flow patterns do not align with sex, indicating the value of personalized over sex-specific approaches in AS characterization, which complement clinical and proteomic findings with the potential to inform treatment strategies.

Patient and Control Characteristics

	Control		Cluster A		Cluster B		p (A vs B)†
	Women	Men	Women	Men	Women	Men
n	19		20		40
Scanner	1.5 T Philips Achieva, five-element cardiac phased-array coil
Flip Angle °	15		15		15
Slice Thickness mm	10		7		7
Repetition Time ms	3.688		3.786 ± 0.112		3.841 ± 0.063
Echo Time ms	2.207		2.300 ± 0.112		2.358 ± 0.063
Venc cm/s	250		600		600
Pixel Spacing mm	1.367		1.208 ± 0.051		1.219 ± 0.001
Sex (male)	10 (53%)		5 (25%)		31 (77.5%)		0.0002*
Age years	62 ± 8		65 ± 10		67 ± 9		0.5249
	59 ± 8	65 ± 8	67 ± 8	59 ± 13	70 ± 5	66 ± 10
BSA m2	1.90 ± 0.23		1.76 ± 0.12		2.07 ± 0.18		< 0.00001*
	1.73±0.13	2.05±0.20	1.72± 0.10	1.90±0.07	1.98±0.14	2.10±0.18
AAo Area mm2	896 ± 235		1033 ± 323		813 ± 235		0.0077*
	887 ± 241	903 ± 242	940 ± 198	1311±479	639 ±215	864 ± 219
LVEF %	64.4 ± 5.2		60.2 ± 7.0		56.1 ± 9.0		0.0436*
	66.3 ± 5.2	62.1± 4.3	60.7 ± 6.9	58.6 ± 7.8	60.6 ±8.1	54.7 ± 8.9
EDVi ml/m2	64.4 ± 11.2		72.0 ± 19.6		83.4 ± 24.8		0.0918
	58.6±9.9	71.9±8.2	69.0±17.1	80.9±25.9	86.3±16.4	87.6±25.3
ESVi ml/m2	23.3 ±6.0		29.4 ± 11.6		37.9 ± 33.2		0.0518
	20.0±5.5	27.4±3.9	27.9±11.1	34.0±13.4	27.9±11.7	40.8±18.0
NT-pro-BNP ng/l	88.7 ± 61.1		356.8 ± 238.2		996.0 ± 1015.2		0.0226*
	120±70	60.2±34.7	408±244	168±58	846±569	1028±1092
Troponin -T ng/l	7.1 ± 3.4		7.8 ± 4.3		13.9 ± 6.8		0.0020*
	7.1 ± 4.3	7.1 ± 2.6	7.6 ± 4.7	8.3 ± 3.5	10.7 ± 7.6	14.5 ± 6.6
Ventricular Scarring			0		9 (22.5%)		0.0179*
BAV			18 (90%)		25 (62.5%)		0.0340*
Valve calcification			1 (5%)		4 (10%)		0.6563

^†Group comparisons with Mann-Whitney U and Fisher's exact test for continuous and count data, respectively. * Indicates statistical significance at α = 0.05.a. 2D PC MRI acquisition plane within the ascending aorta. b. Resampling to uniform voxel count (cohort median) within AAo region results in improved feature stability (here shown: GLRLM Run-Length Non-Uniformity) when features are correlated with AAo size via imaging resolution. Features intrinsically linked to AAo scale irrespective of scale retain correlation under both resampling schemes (not shown). c. Examples of flow profiles in clusters A (blue) and B (yellow). Flow in patients in A is more eccentric, i.e. further from the vessel center. d. Radiomics features with highest weights in Non-Negative Matrix Factorization show cluster differentiation.
a. BSA and Ascending Aortic cross section area, stratified by flow profile cluster and sex. Cluster A contains predominantly women, cluster B mostly men, while women of high BSA and low AAo area are in cluster A, and men with low BSA and high AAo area in cluster B. b. Log2 of protein LFQ intensity of serpinE2 and STAT3, proteins identified in previous studies [3] to be involved in cardiac fibrosis and remodelling.