Cardiovascular Magnetic Resonance Radiomics to Identify Components of the Extracellular Matrix in Dilated Cardiomyopathy

Abstract

Background:

Current cardiovascular magnetic resonance (CMR) sequences cannot discriminate between different myocardial extracellular space (ECS), including collagen, non-collagen, and inflammation. We sought to investigate if CMR radiomics analysis can distinguish between non-collagen and inflammation from collagen in dilated cardiomyopathy (DCM).

Methods:

We identified data from 132 DCM patients scheduled for an invasive septal biopsy who underwent CMR at 3T. CMR imaging protocol included native and post-contrast T₁ mapping and late gadolinium enhancement (LGE). Radiomic features were computed from the mid-septal myocardium, near the biopsy region, on native T₁, extracellular volume (ECV) map, and LGE images. Principal component analysis was used to reduce the number of radiomic features to five principal radiomics. Moreover, a correlation analysis was conducted to identify radiomic features exhibiting a strong correlation (r >0.9) with the five principal radiomics. Biopsy samples were used to quantify ECS, myocardial fibrosis, and inflammation.

Results:

Four histopathological phenotypes were identified as low collagen (n=20), non-collagenous ECS expansion (n=49), mild-to-moderate collagenous ECS expansion (n=42), and severe collagenous ECS expansion (n=21). Non-collagenous expansion was associated with the highest risk of myocardial inflammation (65%). While native T₁ and ECV provided high diagnostic performance in differentiating severe fibrosis (C-statistic: 0.90 and 0.90, respectively), their performance in differentiating between non-collagen and mild-to-moderate collagenous expansion decreased (C-statistic: 0.59 and 0.55, respectively). Integration of ECV principal radiomics provided better discrimination and reclassification between non-collagen and mild-to-moderate collagen (C-statistic: 0.79, net reclassification index (NRI) 0.83; 95% CI 0.45–1.22, p<0.001). There was a similar trend in the addition of native T₁ principal radiomics (C-statistic: 0.75, NRI 0.93; 95%CI 0.56–1.29, p<0.001) and LGE principal radiomics (C-statistic: 0.74, NRI 0.59; 95%CI 0.19–0.98, p=0.004). Five radiomic features per sequence were identified using correlation analysis. They showed a similar improvement in performance for differentiating between non-collagen and mild-to-moderate collagen (native T₁, ECV, LGE C-statistic: 0.75, 0.77, and 0.71, respectively). These improvements remained significant when confined to a single radiomic feature (native T₁, ECV, LGE C-statistic: 0.71, 0.70, and 0.64, respectively).

Conclusions:

Radiomic features extracted from native T₁, ECV, and LGE provide incremental information that improves our capability to discriminate non-collagenous expansion from mild-to-moderate collagen and could be useful for detecting subtle chronic inflammation in DCM patients.

Background

Dilated cardiomyopathy (DCM) is a clinical diagnosis characterized by dilation and systolic dysfunction of the left ventricle (LV) or both ventricles that are not explained by abnormal loading conditions or coronary artery disease.^1,2 In clinical practice, many cases of DCM can be classified as idiopathic after evaluation for excluding secondary causes. However, DCM can result from various etiologies, including inflammation, genetic mutations, and combined factors.² Almost half of DCM cases show histological evidence of myocardial inflammation, which refers to a spectrum of histological phenotypes of the disorder.^3,4 Given that the mortality rates of DCM remain high, and it is a leading cause of heart transplantation,^5,6 more precise phenotyping from histopathology would be welcome.

There have been significant advances in cardiovascular magnetic resonance (CMR), which can provide DCM’s etiologic assessment and risk prediction. Late gadolinium enhancement (LGE) is the gold standard for the assessment of regional myocardial fibrosis and may help predict an anatomic risk for ventricular arrhythmias as well as all-cause mortality in DCM.^7–9 Myocardial T₁ mapping has emerged as a promising imaging technique to non-invasively quantify extracellular volume (ECV)¹⁰ and is an indispensable tool in the risk stratification of DCM patients.^11–13 However, CMR-derived ECV does not necessarily indicate myocardial fibrosis and may result from a non-collagenous matrix, low-grade inflammation/edema, or a combination of inflammation and reactive fibrosis. This is a critically important issue because tailored specific therapies, such as novel strategies designed to mitigate or reverse myofibroblast activation or immunosuppressive therapy, should be provided before developing myocardial fibrosis and could potentially reverse disease progression in DCM. Approaches based on texture analysis and radiomics have been recently proposed to measure new diagnostic and prognostic biomarkers in various cardiovascular diseases.^13–16 However, there are no published reports on the association of CMR radiomics with histopathological findings. Consequently, the objective of this study was to investigate if CMR radiomics analysis can distinguish between non-collagen and inflammation from mild-to-moderate collagen in recent-onset DCM patients.

Methods

Anonymized data and materials have been made publicly available at the Radiomics_Histology_PCA and can be accessed at https://github.com/HMS-CardiacMR.

Study population

In this retrospective study, we identified 132 Asian patients with recent-onset DCM who underwent CMR, invasive coronary angiography, and endomyocardial biopsy (EMB) between September 2015 and January 2022 (Figure S1). EMB was performed for initial evaluation to exclude other etiologies.¹⁷ The diagnosis of DCM was based on medical history, physical examination, electrocardiography, echocardiography, CMR, coronary angiography, and EMB. Exclusion criteria included: 1) ischemic etiology (defined as the presence of an epicardial coronary artery diameter stenosis > 50%, a history of myocardial infarction, or a subendocardial-based LGE pattern in a coronary distribution), 2) clinically diagnosed myocarditis, hypertrophic and infiltrative cardiomyopathies, and more than moderate valvular heart diseases 3) persistent atrial fibrillation, or 4) any contraindications for CMR. The study was carried out with Mie University Hospital Institutional Review Board approval (reference number: H2020–113) which waived written informed consent for the retrospective study. All anonymized DICOM images were transferred to Beth Israel Deaconess Medical Center in Boston, MA for further analysis.

Endomyocardial biopsy

EMB samples were taken from the interventricular septum using a 5.4 Fr bioptome (ab medica sas, Méry-sur-Cher, France) via the right jugular or femoral vein, and at least two or three samples per patient were then submitted for light microscopic examination. All biopsy specimens were immediately fixed in a 10% buffered formalin, embedded in paraffin wax, sectioned at 3 μm, and stained with picrosirius red to determine histological collagen volume fraction (CVF) or with hematoxylin-eosin for analysis of the ECS component. In some cases, Congo red and Direct fast scarlet 4BS staining were added to exclude amyloid deposits. Five to seven high-power (200×) magnification digital images were acquired per patient for image analysis (BZ-X710; Keyence, Osaka, Japan). Colored images with picrosirius red staining were split into multiple color channels. Using an image with the best separation of myocardial fibrosis, CVF was calculated as the percentage of collagen area divided by total myocardial area.^10,18,19 Image processing with a freehand selection tool was performed to exclude the subendocardial area. The fibrosis was calculated as the average fibrosis burden of all histological slides in a patient.²⁰ Similarly, in a hematoxylin-eosin-stained section, a color-based calculation algorithm derived from bright pink staining of myocyte and stained purple of nucleus was applied to yield percentage areas of ECS (ImageJ version 1.50b, National Institution of Health, Bethesda, MD, USA) (S.N., with 15 years of experience and K.I.Y., with more than 30 years of experience).¹⁰ Non-collagenous ECS was calculated as total ECS minus CVF (Figure 1A). Immunohistochemistry analysis was performed to identify the type of inflammatory infiltrate using anti-CD3 (IR50361–2, Agilent) for T-lymphocytes, and anti-CD68 (M081401–2, Agilent, Santa Clara, CA, USA) for macrophages. Subjects were classified into four histopathological phenotypes: low CVF (CVF <5%); non-collagenous ECS expansion of ≥15% (5%≤ CVF <20%); mild-to-moderate collagenous ECS expansion (non-collagenous expansion <15% and 5%≤ CVF <20%); and severe collagenous ECS expansion (CVF of 20% or larger) (Figure 1B). Myocardial inflammation was considered positive if ≥14 leukocytes/mm², including ≤4 monocytes/mm² with the presence of CD3+ T-lymphocytes ≥7 cells/mm² was detected in EMB specimens according to the ESC criteria.²¹

Figure 1: Measurement of collagen volume fraction and extracellular space and histopathological phenotyping.

(A) Images are hematoxylin-eosin-stained and picrosirius red-stained tissue sections. By using ImageJ software (National Institution of Health, Bethesda, MD, USA), high-power magnification digital images (200×) were used for semiautomated image analysis to assess ECS (top row) and CVF (bottom row), which were calculated as the percentage of ECS area (black area) and collagen area (light blue area), respectively, divided by total myocardial area, excluding subendocardial areas. Non-collagenous ECS was calculated as total ECS minus collagen area. ECS, CVF, and non-collagenous ECS, in this case, correspond to 24%, 6%, and 18%, respectively. (B) Subjects were classified into four histopathological phenotypes: low CVF (CVF <5%); mild-to-moderate fibrosis (5%≤ CVF <20%) with non-collagenous ECS expansion ≥15%; mild-to-moderate fibrosis (5%≤ CVF <20%) without non-collagenous ECS expansion≥15%; and severe fibrosis (20%≤ CVF)

CVF; collagen volume fraction, ECS; extracellular space

Image acquisition and image analysis

All images were acquired with a 3T CMR scanner (Ingenia 3T, Philips Healthcare, Best, The Netherlands) equipped with a commercial torso coil. The CMR protocol included cine, native and post-contrast T₁ mapping by modified Look-Locker inversion recovery (MOLLI) sequence, and LGE. The LGE images were collected 10–15 minutes after 0.15mmol/kg bolus injection of Gd-DOTA (Magnescope, Guerbet, Japan). To assess LV myocardial function and mass, 10 to 12 short-axis stack cine images and 2-, 3-, and 4-chamber long-axis images were acquired using a cine balanced steady-state free precession sequence (slice thickness=10 mm, 50 ms temporal resolution). Native T₁ was performed at 3 different slice locations covering the whole LV from base to apex using a steady-state free procession, single-breath-hold 3-3-5 MOLLI (TR/TE=2.6/1.1 ms, flip angle=35°, FOV=300×330 mm², acquisition matrix=176×141, reconstruction matrix=288×288, slice thickness=10 mm, SENSE factor=2, diastolic acquisition). LV short-axis 3D LGE images were acquired with an inversion recovery gradient-echo imaging sequence using the following imaging parameters: TR/TE=4.7/2.3 ms, flip angle =15°, FOV= 38×34×5 cm³, acquisition matrix=240×192×5, acquisition thickness=10 mm, reconstruction matrix=384×384×10, reconstructed thickness=5 mm, and SENSE factor 3. Post-contrast 3-3-5 MOLLI T₁ mapping was performed exactly as pre-contrast T₁ mapping 15 minutes after administration of the gadolinium contrast.

CMR images were analyzed on a commercially available post-processing workstation cvi42 version 5.11.1 (Circle Cardiovascular Imaging Inc. Calgary, Alberta, Canada) and assessed by the consensus of two experienced readers (S.N., with 10 years of experience, and M.I., with more than 20 years of experience). On cine CMR images, epicardial and endocardial LV borders at end systole and end diastole were manually traced from contiguous short-axis cine images covering the LV apex to the mitral valve plane to calculate LV end-diastolic volume and end-systolic volume, and LV ejection fraction (LVEF). LV mass was calculated as the sum of the myocardial volume multiplied by the specific gravity (1.05g/mL) of myocardial tissue. T₁ maps were estimated by voxel-wise curve fitting of the signal using a 2-parameter fit model. T₁ values of LV myocardium and blood pool were similarly determined before and after contrast injection, and ECV was computed, as previously described.²² Native T₁ and ECV values were measured conservatively within the inferior septal myocardium in a mid-ventricular short-axis slice, corresponding to the biopsy region. For Native T₁ and ECV, we did not exclude areas of scar, if present on LGE. On LGE images, the presence and location of LGE were visually assessed. LGE volume was measured using the full width at half maximum method (S.N., with 10 years of experience).

Radiomic Analysis

For radiomic analysis, a single reader (S.N.) placed a region of interest (ROI) on the septal myocardium at the mid-ventricular level near the EMB region, of an average size 780 mm² (23×30 pixels), using each image acquired during the diastolic phase. Six months later, a second reader (A.A.) repeated segmentation in 65 patients blinded to the first reader for inter-observer reproducibility assessment of selected radiomic features. We normalized the ROI spatial resolution to 1×1 mm² using bilinear interpolation. Subsequently, we normalized the intensity values within the ROI to a range of 0 to 1. This normalization was achieved by applying the following equation to each pixel p within the ROI: $I_p=\frac{I_p-{\min }_{r\in R}{I}_r}{{\max }_{r\in R}{I}_r-{\min }_{r\in R}{I}_r}$, where $I_p$ represents the intensity of a specific pixel p, ${\min }_{r\in R}{I}_r$ denotes the minimum intensity value among all pixels within the ROI, and ${\max }_{r\in R}{I}_r$ denotes the maximum intensity value among all pixels within the ROI. The set R encompasses all pixels within the ROI. Radiomic features were then calculated from native T₁, ECV, and LGE images using the open-source library Pyradiomics (version 3.0.1),²³ and 1023 first-order and texture features were extracted per sequence (e.g., energy or entropy of myocardial intensity histogram, gray-level run-length matrix, gray-level co-occurrence matrix, and local binary patterns) computed from each ROI using 9 types of filters (gradient, logarithmic, squaring, square-root, exponential, and 4 wavelet-transform filters). Shape features were excluded because the ROI near the EMB region contained no meaningful information. Due to the redundancy of many radiomic features, highly correlated features with a correlation coefficient >0.8 were removed. To further reduce the dimensionality of the radiomic features, principal component analysis (PCA) was performed to extract the components representing the top variations of the radiomic features per sequence. The top five components were selected to have a ratio of 1 predictor to every 10 non-collagenous patients with at least 50% explained variance, referred to as principal radiomics. It is worth noting that the feature selection process was unsupervised and independent of the output, as it depends only on the feature variance (Figure 2).

Figure 2: Proposed radiomics-driven feature model using native T₁, ECV, and LGE.

For the radiomic feature analyses, a large region of interest was placed on the inferior septal myocardium at the mid-ventricular level of each image, corresponding to the biopsy region. A total of 1023 first-order and textural features were extracted per sequence, and principal component analysis was performed to extract the components representing the top variations of the radiomic features per sequence. For model development, we selected the top 5 components (principal radiomics) based on the number of phenotypes in the dataset.

DCM; dilated cardiomyopathy, ECS; extracellular space, ECV; extracellular volume, LGE; late gadolinium enhancement

Interpretation of principal radiomics

The principal radiomics represent orthogonal linear transformations of several dozen radiomic features calculated from the images. This can lead to challenges in their clinical interpretation. In a secondary analysis, we sought to investigate if any individual or subset of radiomic features provides similar model performance as the entire radiomic features. To accomplish this goal, we identified radiomic features exhibiting high correlation (r >0.9) with each sequence’s five principal components. This approach eliminates any leakage of information from the outcome of interest to feature selection. We selected the top five radiomic features from this analysis and re-evaluated the model performance by adding these new radiomic features instead of the principal radiomics. Furthermore, to determine a representative radiomic feature, we used the area under the receiver operating characteristic (ROC) curve to identify one feature from the five selected radiomic features.

Statistical analysis

Statistical analyses were performed using SPSS version 25 software (IBM Inc., Chicago, IL, USA) and R version 3.2.3 (R Project for Statistical Computing). Continuous variables are expressed as mean ± standard deviation (SD) or median [quartiles] as appropriate. Categorical variables were reported as counts and percentages and compared using a chi-square or Fisher`s Exact test if at least one of the expected cell counts was below 5. One-way ANOVA with Bonferroni adjustment or the Kruskal-Wallis test followed by Dunn’s procedure, if the data did not follow a normal distribution, was applied for multiple comparisons after the 4 group comparisons. Pearson/Spearman correlation coefficient was used to examine possible relationships among LV function, histological CVF, ECS, and CMR T₁ mapping findings. For each sequence, we built 4 logistic regression models using: 1) native T₁ values, ECV, or presence of LGE, 2) native T₁ values, ECV, or presence of LGE + principal radiomics, 3) native T₁ values, ECV, or presence of LGE + five radiomic features, and 4) T₁ values, ECV, or presence of LGE + single radiomic features. The area under the ROC curve was calculated and compared with a DeLong test for all predictive tests for phenotyping histological category. Reclassification of patients was determined using net reclassification improvement analysis for phenotyping non-collagenous expansion group and obtained by adding CMR radiomics to clinical CMR assessments. For inter-observer reproducibility of selected radiomic features, the intraclass correlation coefficient (ICC) was calculated. ICC estimates and their 95% confidence intervals were calculated using Pingouin statistical package version 0.5.3 based on a 2-way mixed-effects model with absolute agreement. Based on ICCs, agreement was defined as fair (0.40 to 0.59), good (0.60 to 0.74), and excellent (≥ 0.75). All tests were 2-sided, and p-values <0.05 were considered significant.

Results

Patient characteristics

The baseline clinical characteristics of 132 DCM patients (54 ± 15 years, 38 [29%] females) are separated into four groups according to histological classification and shown in Table 1. At the time of the CMR scan, 92% of patients were considered New York Heart Association (NYHA) functional class ≥ II. No patients received device implantation before CMR. Overall, 36% of patients had a history of hypertension, 20% had diabetes, 14% had moderate excess alcohol consumption, and 5% had received prior chemotherapy. The usage of angiotensin-converting enzyme inhibitor/angiotensin receptor blockers, beta-blockers, and mineralocorticoid antagonists before optimal medical therapy were comparable among the 4 groups. There was a significant difference in NYHA functional class, brain natriuretic peptide (BNP) levels, or renal function levels among the four groups. Patients with severe collagenous expansion also tended to have lower systolic blood pressure and longer duration of heart failure. Patients with non-collagenous and mild-to-moderate collagenous expansion were balanced in demographics and disease characteristics. The median duration between CMR and biopsy was 6 days.

Table 1.

Comparison of baseline clinical and laboratory findings among 4 different histopathological phenotypes

	All patients	Low CVF	Non-collagen	Mild-to-moderate collagen	Severe collagen	p-Value
	(n = 132)	(n = 20)	(n = 49)	(n = 42)	(n = 21)
Clinical characteristics
Age, yrs	54 ± 15	57 ± 15	51 ± 16	55 ± 15	59 ± 14	0.23
Female gender, %	38 (29)	6 (30)	17 (35)	11 (26)	4 (19)	0.57
Body surface area, m2	1.69 ± 0.23	1.74 ± 0.25	1.63 ± 0.22	1.73 ± 0.24	1.73 ± 0.22	0.10
Systolic blood pressure, mmHg	119 ± 22	113 ± 21	122 ± 25	123 ± 22	110 ± 15	0.07
NYHA functional status						0.002
Class I, %	10 (8)	4 (20)	4 (8)	1 (2)	1 (5)
Class II, %	77 (58)	13 (65)	31 (63)	25 (60)	8 (38)
Class III, %	40 (30)	3 (15)	14 (29)	15 (36)	8 (38)
Class IV, %	5 (4)	0	0	1 (2)	4 (19)
Duration of heart failure, months	2.3 ± 1.4	1.9 ± 1.0	2.1 ± 1.2	2.5 ± 1.5	2.9 ± 1.7	0.08
Time interval between CMR and biopsy, days	6 (2–13)	5 (2–18)	6 (2–17)	6 (2–12)	5 (2–32)	0.80
Comorbidities
Hypertension, %	48 (36)	8 (40)	15 (31)	18 (43)	7 (33)	0.64
Diabetes mellitus, %	26 (20)	5 (25)	9 (18)	8 (20)	4 (19)	0.94
Moderate alcohol excess, %	18 (14)	2 (10)	6 (12)	9 (21)	1 (5)	0.26
Previous chemotherapy, %	7 (5)	1 (5)	4 (8)	2 (5)	0	0.39
Medication
ACE inhibitor and/or ARB, %	107 (81)	14 (70)	43 (88)	32 (76)	18 (86)	0.26
Beta blockers, %	69 (52)	11 (55)	25 (51)	23 (55)	10 (48)	0.94
MRA, %	41 (31)	7 (35)	13 (27)	16 (38)	5 (24)	0.55
Diuretic agents, %	86 (65)	12 (60)	28 (57)	31 (74)	15 (71)	0.33
Laboratory data
BNP, pg/ml	124.9 (52.5–334.8)	88.8 (29.0–145.5)	124.9 (53.5–347.5)	90.2 (47.1–227.8)	357.3 (103.6–643.4)	0.02
BUN, mg/dl	17.4 ± 5.0	18.0 ± 5.9	16.3 ± 4.7	16.8 ± 4.5	20.9 ± 4.6	0.003
Creatinine, mg/dl	0.96 ± 0.25	1.01 ± 0.29	0.93 ± 0.27	0.92 ± 0.24	1.06 ± 0.17	0.13
eGFR, ml/min/1.73 m2	68.8 ± 21.8	68.7 ± 24.3	71.8 ± 25.0	71.6 ± 18.5	56.2 ± 12.4	0.03
Electrocardiographic data
LBBB, %	4 (3)	0	3 (6)	1 (2)	0	0.28
QRS duration, msec	99 ± 13	101 ± 14	99 ± 12	100 ± 15	97 ± 9	0.73

ACEi/ARB=angiotensin-converting enzyme inhibitors/ angiotensin-receptor blockers; BNP=brain natriuretic peptide; BUN=blood urea nitrogen; CMR=cardiovascular magnetic resonance; eGFR=estimated glomerular filtration rate; LBBB=left bundle branch block; MRA=mineralocorticoid receptor antagonists; NYHA=New York Heart Association;

Clinical CMR and histopathological findings

CMR and histological characteristics are summarized in Table 2. The mean LVEF was 31%, and 78 patients (59%) had a nonischemic LGE. There was a significant difference in LV cavity size, LGE prevalence, and extent, native T₁ or ECV among the four groups. The patients with low CVF had the lowest prevalence and extent of LGE, native T₁, and ECV, while those with non-collagenous expansion had the smallest LV cavity size of the other groups. The patients with severe fibrosis showed the largest LV cavities and highest in the rate and extent of LGE, native T₁, and ECV values. Native T₁ and ECV values across all patients are shown in Figure 3. The mean native T₁ and ECV in the non-collagenous expansion group were 1356 ± 73 ms and 33.3 ± 4.8%, respectively, while the native T₁ and ECV were 1335 ± 71 ms and 32.3 ± 4.0%, respectively, in mild-to-moderate collagenous expansion group, indicating a substantial overlap between the groups.

Table 2.

CMR and histological characteristics among 4 different histopathological phenotypes

	All patients	Low CVF	Non-collagen	Mild-to-moderate collagen	Severe collagen	p-Value
	(n = 132)	(n = 20)	(n = 49)	(n = 42)	(n = 21)
CMR findings
LV ejection fraction, %	31.3 ± 12.2	30.1 ± 12.1	33.8 ± 12.2	30.9 ± 12.7	27.5 ± 10.5	0.23
LV end-diastolic volume, ml	230.2 ± 78.4	237.2 ± 81.2	204.6 ± 77.4*	237.5 ± 72.9	268.9 ± 73.4	0.01
LV end-diastolic volume index, ml/m2	136.6 ± 44.7	138.3 ± 51.4	125.9 ± 44.5	137.8 ± 38.3	158.0 ± 45.0	0.05
LV end-systolic volume, ml	163.3 ± 76.8	172.0 ± 81.4	138.8 ± 72.7	169.4 ± 72.7	199.9 ± 76.0	0.01
LV end-systolic volume index, ml/m2	98.5 ± 45.8	101.2 ± 51.9	89.1 ± 45.8	98.5 ± 40.7	118.0 ± 46.4	0.11
LV mass, g	126.5 ± 43.4	140.1 ± 59.2	118.2 ± 40.9	127.1 ± 36.2	131.8 ± 44.3	0.26
LV mass index, g/m2	74.7 ± 23.7	78.7 ± 27.4	72.6 ± 22.9	74.3 ± 21.4	76.7 ± 27.4	0.78
Presence of LGE, %	78 (59)	8 (40)	26 (53)	26 (62)	18 (86)	0.01
Extent of LGE, %	2.5 (1.9–5.5)	1.3 (1.0–4.4)	3.2 (2.4–4.8)*	1.9 (1.6–2.1)	7.6 (6.2–8.6)	<0.001
Hematocrit, %	42.3 ± 8.9	43.6 ± 7.2	41.4 ± 9.3	43.0 ± 9.6	41.5 ± 8.0	0.72
Heart rate, bpm	73 ± 13	76 ± 15	71 ± 13	74 ± 13	74 ± 12	0.48
Septal native T1, ms	1348 ± 71	1304 ± 43	1356 ± 73	1335 ± 71	1395 ± 57	<0.001
Septal ECV, %	33.4 ± 5.0	30.7 ± 4.0	33.3 ± 4.8	32.3 ± 4.0	38.5 ± 4.9	<0.001
Histological findings
Extracellular space, %	27.6 ± 7.3	18.2 ± 2.5	31.3 ± 4.8*	23.8 ± 4.0	35.5 ± 6.3	<0.001
Collagen volume fraction, %	12.6 ± 7.3	4.2 ± 1.0	9.9 ± 4.4*	13.7 ± 3.3	24.8 ± 5.9	<0.001
Non-collagenous space, %	15.0 ± 6.3	14.0 ± 2.1	21.5 ± 4.3*	10.2 ± 2.8	10.7 ± 4.5	<0.001
Presence of inflammation, %	52 (39)	3 (15)	32 (65)*	10 (24)	7 (33)	<0.001

^*Non-collagen vs. Mild-to-moderate collagen p <0.05

CMR=cardiovascular magnetic resonance; ECV=extracellular volume; LGE=late gadolinium enhancement; LV=left ventricular

Figure 3: Comparisons of native T₁ and ECV among 4 different histopathological phenotypes.

Native T₁ values provide high diagnostic concordance in predicting histological phenotype severity associated with myocardial fibrosis. However, native T₁ values only mildly increased and substantially overlapped between the non-collagenous and collagenous expansion. There is a similar trend in ECV values. Reference values at our institution; native T1:1314±29 ms, ECV: 26±4%

ECS; extracellular space, ECV; extracellular volume

Histological CVF and ECS were 12.6 ± 7.3% and 27.6 ± 7.3%, respectively. The patients with severe collagenous expansion had the highest in both CVF and ECS. The non-collagenous expansion group had significantly lower CVF, higher ECS, and non-collagenous space compared with the mild-to-moderate collagenous expansion group. Also, myocardial inflammation was observed in 39% of patients, and 71% (30/42) of patients with non-collagenous expansion showed the highest risk of inflammation than the other groups. Native T₁ and ECV similarly and moderately correlated with CVF (r=0.49 and r=0.52, respectively; p<0.001 for both). ECV, rather than native T₁, correlated well with ECS (r=0.56 and r=0.48, respectively; p<0.001 for both). The C-statistics of native T₁ and ECV mapping in differentiating severe fibrosis were 0.90 (95%CI 0.80–1.00) and 0.90 (95%CI 0.80–0.99). In contrast, the performances of native T₁ and ECV mapping in differentiation between non-collagenous and mild-to-moderate collagenous expansion decreased to the C-statistics of 0.59 (95%CI 0.47–0.71) and 0.55 (95%CI 0.43–0.67), respectively.

Principal radiomics for differentiation between non-collagen and mild-to-moderate collagen

The C-statistics for the principal radiomics models in predicting non-collagen deposition were 0.74 (95%CI 0.63–0.85) in native T₁, 0.78 (95%CI 0.68–0.88) in ECV, and 0.72 (95%CI 0.62–0.83) in LGE, respectively. When ECV principal radiomics were combined with clinical assessment of ECV, we observed a greater C-statistic from 0.55 (95%CI 0.43–0.67) to 0.79 (95%CI 0.69–0.89) (DeLong p-value; 0.001). Overall, integration of ECV principal radiomics to ECV value provided improvement in tissue phenotyping (net reclassification index (NRI) 0.83; 95% CI 0.45–1.22, p<0.001). There was a similar trend in the addition of native T₁ principal radiomics to native T₁ (C-statistic=0.75 [95% CI 0.65–0.86] vs. 0.59 [95%CI 0.47–0.71], DeLong p-value; 0.02, NRI 0.93; 95%CI 0.56–1.29, p<0.001) and LGE principal radiomics to LGE (C-statistic=0.74 [95% CI 0.64–0.84] vs. 0.54 [95%CI 0.44–0.65], DeLong p-value; 0.003, NRI 0.59; 95%CI 0.19–0.98, p=0.004) (Figure 4).

Figure 4: ROC curves of principal radiomics, pre-selected five radiomic features, and one radiomic feature for discriminating the non-collagen and mild-to-moderate collagen.

Receiver operating characteristic (ROC) curve and corresponding area under the curve (AUC) for differentiation of non-collagenous and collagenous deposition

ECV; extracellular volume, LGE; late gadolinium enhancement

Identification of radiomic features and their performance for differentiation between non-collagen and mild-to-moderate collagen

Table 3 depicts radiomic features which strongly correlate (r>0.9) with each of the top 5 principal radiomics, and Table S1 shows the diagnostic performance of these radiomic features for the diagnosis of non-collagen deposition. Among the 8 radiomic features highly correlated with the 5 principal radiomics of the native T₁ map, the contrast of the myocardial neighboring gray-tone difference matrix (NGTDM) only showed a significant discriminative value between non-collagenous and mild-to-moderate collagenous expansion groups (C-statistic=0.68, 95%CI 0.56–0.79), while the busyness derived from NGTDM and interquartile range from first-order features allowed the best discrimination in ECV radiomic features and LGE radiomic features (C-statistic=0.74 [95%CI 0.63–0.84] and 0.64 [95%CI 0.53–0.76], respectively). All these radiomic features were significantly lower in the non-collagenous expansion group than the mild-to-moderate collagenous expansion group (Figure S2), indicating a more homogeneous or uniform texture in the non-collagenous expansion group. Figure 5 depicts representative cases of non-collagenous expansion and collagenous expansion groups.

Table 3.

Radiomic features highly correlated with principal radiomics (r>0.9)

	Explained variance ratio	Radiomic features
Native T1 Radiomics
Principal radiomics 1	19.15%	NGTDM-Contrast
Principal radiomics 2	13.13%	GLCM-Imc1 and GLCM-Correlation
Principal radiomics 3	9.13%	NGTDM-Coarseness and First Order-Variance
Principal radiomics 4	7.95%	First Order-Uniformity
Principal radiomics 5	5.30%	GLCM-Cluster Tendency and NGTDM-Strength
ECV Radiomics
Principal radiomics 1	20.28%	NGTDM-Busyness
Principal radiomics 2	11.62%	GLCM-Maximum Probability
Principal radiomics 3	9.10%	GLCM-Imc1
Principal radiomics 4	7.49%	First Order-Kurtosis
Principal radiomics 5	5.66%	GLRLM-Long Run Low Gray Level Emphasis
LGE Radiomics
Principal radiomics 1	20.14%	NGTDM-Contrast and First Order-Interquartile Range
Principal radiomics 2	13.51%	First Order-Mean
Principal radiomics 3	9.61%	NGTDM-Strength and GLCM-Maximum Probability
Principal radiomics 4	6.11%	First Order-10th Percentile and First Order-90th Percentile
Principal radiomics 5	5.43%	First Order-Entropy

ECV=extracellular volume; LGE=late gadolinium enhancement; GLCM=gray-level co-occurrence matrix; GLRLM= gray-level run length matrix; Imc1=informational measure of correlation 1; NGTDM= neighbouring gray tone difference matrix

Figure 5: Representative cases of non-collagenous and collagenous expansion with similar extracellular space.

(A) Hematoxylin-eosin and picrosirius red-stained samples showed ECS expansion of 33% but less fibrosis (CVF=9%). Immunohistochemistry of this case confirmed myocardial inflammation. CMR images showed a relatively smaller LV cavity size without LGE. (B) The ECS and CVF examples corresponded to 31% and 18%, which revealed the difference in CVF despite similar ECS expansion, and no inflammation was observed. Native T₁, ECV, and LGE texture features of case B seem to be heterogeneous, comparing those of case A. There was a small amount of LGE in the inferior wall.

CVF; collagen volume fraction, ECS; extracellular space, ECV; extracellular volume, LGE; late gadolinium enhancement, LV; left ventricular

When using ECV radiomics with the pre-selected 5 ECV radiomic features with high diagnostic performance, the C-statistic increased from 0.55 (95%CI 0.43–0.67) to 0.77 (95%CI 0.66–0.87) (DeLong p-value; 0.002). Overall, integration of the 5 radiomic features to ECV value provided improvement in risk stratification (net reclassification index (NRI) 0.67; 95% CI 0.28–1.06, p<0.001). There was a similar trend in the addition of pre-selected 5 native T₁ and LGE radiomic features to native T₁ and LGE (C-statistic=0.75 [95% CI 0.65–0.85] vs. 0.59 [95%CI 0.47–0.71], DeLong p-value; 0.01, NRI 0.77; 95%CI 0.40–1.13, p<0.001, C-statistic=0.71 [95%CI 0.61–0.82] vs. 0.54 [95%CI 0.43–0.66], DeLong p-value; 0.007, NRI 0.52; 95%CI 0.13–0.91, p=0.009) (Figure 4). When confined to a single radiomic feature with the highest diagnostic performance, these improvements remained significant (native T₁, ECV, LGE C-statistic: 0.71 [95%CI 0.60–0.82], 0.70 [95%CI 0.59–0.81], and 0.64 [95%CI 0.53–0.76], respectively). Table S2 shows the enhanced diagnosis of non-collagen according to the number of radiomic features.

Inter-observer variability of radiomic features

For native T₁, the inter-observer reproducibility was good to excellent for all highly correlated features, with GLCM-Imc1 being the most reproducible feature (ICC=0.81) (Table S3). For ECV, the First Order-Kurtosis and GLRLM-Long Run Low Gray Level Emphasis showed fair reproducibility, while the other three features showed good to excellent reproducibility. GLCM-Maximum Probability was the most reproducible feature (ICC=0.85) (Table S4). All but one of the LGE features (First Order-10th Percentile, ICC=0.70) had excellent reproducibility, with NGTDM-Contrast being the most reproducible feature (ICC=0.94) (Table S5). Nearly all radiomic features that highly correlated with principal radiomics had good to excellent reproducibility.

Discussion

In this retrospective histological validation of CMR radiomics of 132 recent-onset DCM patients, we demonstrate that 1) clinical native T₁ and ECV mapping provide high diagnostic concordance in predicting histological phenotype severity associated with myocardial fibrosis, 2) absolute values of native T₁ and ECV, however, show limited diagnostic accuracy in detecting mild-to-moderate myocardial tissue alteration and differentiating non-collagenous ECS from mild-to-moderate collagen, 3) integration of CMR radiomics with clinical assessment provides better ability to predict the non-collagenous ECS expansion, often accompanied by inflammation, from mild-to-moderate collagen.

DCM is an “umbrella” term describing the final common phenotype of various etiologies after excluding secondary cardiomyopathy such as myocarditis, ischemic heart disease, or valvular disease. Considering myocardial inflammation typically precedes the development of fibrosis at the site of tissue damage and a pivotal process associated with poor prognosis,^24–26 rather than focusing on indirect measurement such as LV dysfunction and dilatation, a better understanding of myocardial histology may be of great importance and inform a faster pathway for development of precision medicine. Our imaging analysis is a first attempt towards personalized management of DCM patients using EMB as the reference standard. Patients with non-collagenous ECS expansion and less collagen demonstrated a trend toward shorter duration of heart failure as well as smaller LV cavities, suggesting early myocardial damage at this stage than in those with diffuse collagen deposition. Conversely, non-collagenous expansion group had a trend toward a higher extent of LGE although LGE was common in both non-collagenous and mild-to-moderate collagenous groups, which might be explained by subtle ongoing edema/inflammation followed by scar-shrinking over time accompanied by reduced spatial extent. Actually, 65% (32/49) of patients with non-collagenous expansion showed immunohistological evidence of chronic inflammation based on ESC criteria, which was the highest than the other groups. Unfortunately, native T₁ and ECV values were only mildly increased and overlapped in both groups, which corroborates the results of previous studies comparing T₁ indices and histology, indicating differential diagnosis challenges. CMR radiomics enable more precise myocardial tissue characterization and have a promising potential to unlock and demystify many of the blind spots of the current diagnosis, management, and risk stratification of DCM. The understanding of radiomics based on histology and the unraveling of the radiomic features of each variant should dictate different therapeutic strategies.

Despite the promise of radiomics for uncovering new insights from medical images,²⁷ there has been growing concerns regarding its translation to clinical practice.²⁸ Recent guidelines have been proposed to address study design, statistical analysis, and strategies for mitigating bias.^29,30 Two prominent techniques in machine learning (ML) are supervised and unsupervised learning.³¹ The conventional radiomic analysis approach involves extracting pre-engineered features from original and/or filtered images and using supervised algorithms to identify the most predictive radiomic features for the outcome of interest. This approach provides transparency and emphasizes predictive features, but needs to be validated on a separate dataset not used during training to ensure robustness and generalizability.³² This process entails the optimization of the ML algorithm during training and the leakage of the outcome information during the selection process of the radiomic features. The resulting radiomic features are usually biased by the optimization and feature selection processes. We used PCA for feature selection, an unsupervised method for dimensionality reduction independent of the output/outcome. This approach ensures that the selected features are not biased by optimization algorithms or the specified output, thus enhancing their reproducibility. However, this may come at the cost of sacrificing some interpretability and overlooking data relationships. To address this, we conducted a correlation analysis between principal components and radiomic features, providing insights into interpretability.

The goal of our study was to identify properties or characteristics that exhibit significant differences among patients with mild-to-moderate collagenous vs. non-collagenous matrix, often accompanied by inflammation. Thus, radiomic features with high variance may provide valuable information in this differentiation. PCA helps capture essential information about various characteristics such as first-order statistics (minimum, maximum, entropy, etc.) and changes in tissue intensities (heterogeneity), essential in distinguishing mild-to-moderate collagenous from non-collagenous matrix. To gain insight into clinical meaning, we identified radiomic features that highly correlated (r >0.9) with the extracted 5 principal radiomics from each sequence and assessed their diagnostic performance. The NGTDM quantifies the difference between a gray value and the average gray value of its neighboring pixels. Low NGTDM contrast and busyness values indicate a more homogeneous or uniform tissue texture.³³ While these features have demonstrated the ability to differentiate tumor tissue from normal tissue in head and neck cancer,^34,35 their underlying histopathological significance in cardiac imaging is unknown. Our research has unveiled a link between a more uniform or homogeneous texture and non-collagenous matrix, typically accompanied by myocardial inflammation. Conversely, increased heterogeneity is linked with collagen deposition, particularly diffuse fibrosis. Similar to clinical assessments, first-order features are also commonly employed as biomarkers, and a low interquartile range may indicate myocardial tissue homogeneity.

Currently, CMR can detect myocardial edema/inflammation suggestive of acute myocardial infarction, active myocarditis, or sarcoidosis. In a study of 57 DCM patients by Spieker, et al., clinical T₂ mapping approach detected subtle inflammation of DCM patients.³⁶ However, in a sub-group analysis of 64 patients with T₂ mapping (Supplemental Materials), we found no significant difference in T₂ values between patients with and without histological myocardial inflammation. Myocardial T₂ is theoretically more sensitive to changes in myocardial water content secondary to myocardial edema/inflammation. However, the current T₂ mapping technique may not have the necessary sensitivity needed to detect changes in myocardial T₂ associated with a subtle change of non-collagenous ECS expansion or inflammation/edema seen in DCM patients. Thus, in routine clinical practice, many DCM patients have borderline normal T₂ values, which potentially leads to false-negative exclusion of inflammation. Further studies are needed to confirm the utility of clinical T₂ mapping and the potential of T₂ mapping-based radiomics in DCM patients because T₂ mapping was unavailable for all patients in this study.

Study limitations

Our study has several limitations. This is a retrospective single-center modest size study and images were collected using a single vendor. All data were collected at 3T, and our population was all Asian. Further studies are warranted to investigate the generalizability of our findings. Right-sided biopsies might be associated with an increased risk of sampling error to define cardiomyopathic processes in the LV. However, a previous study demonstrated only a limited value of LV biopsies. No real-time polymerase chain reaction (PCR) was performed.

Conclusion

CMR radiomics has the potential to discriminate non-collagenous ECS expansion from mild-to-moderate collagen and may be useful for detecting myocardial inflammation in recent-onset DCM patients, thereby avoiding invasive EMB. Further studies are warranted to examine the utility of this CMR radiomics in the selection and risk stratification of DCM patients who are more likely to have subtle chronic inflammation.

Clinical Perspective.

What is new?

• Clinical native T₁ and ECV mapping show limited diagnostic accuracy in detecting mild-to-moderate myocardial tissue alteration and differentiating non-collagenous extracellular space (ECS) from mild-to-moderate collagen.

• Radiomic features extracted from native T₁, ECV, and LGE provide incremental information that improves our capability to discriminate non-collagenous expansion, often accompanied by myocardial inflammation, from mild-to-moderate collagen.

What are the clinical implications?

• CMR radiomics has the potential to discriminate non-collagenous ECS expansion from mild-to-moderate collagen and may be useful for detecting subtle chronic inflammation, thereby providing more precise diagnosis and tailored treatment strategies in recent-onset DCM patients.

Non-standard Abbreviations and Acronyms

CVF

collagen volume fraction

CMR

cardiovascular magnetic resonance

DCM

dilated cardiomyopathy

EMB

endomyocardial biopsy

ECS

extracellular space

ECV

extracellular volume

EF

ejection fraction

LGE

late gadolinium enhancement

LV

left ventricular/ventricle

ML

machine learning

NGDTM

neighboring gray-tone difference matrix

NRI

net reclassification index

PCA

principal component analysis