The Importance of Spatial Resolution in Detecting Chronic Myocardial Iron Remnants Following Hemorrhagic Myocardial Infarction with T2* MRI

Abstract

Purpose:

Intramyocardial hemorrhage following revascularized myocardial infarction is associated with increased mortality and morbidity. T2* cardiac MRI is the noninvasive imaging standard for visualizing the chronic iron deposits from hemorrhagic myocardial infarction (hMI). However, the impact of spatial resolution on detection performance of these iron deposits remains unclear.

Methods:

We investigated the influence of spatial resolution on the detection of iron deposits following hMI using (i) digital phantom simulations across voxel sizes (0.4–2.0 mm² in-plane; 2–8 mm slice thickness) at multiple SNR levels, (ii) ex-vivo swine hearts (n=10) with 8-week hMI. Ex-vivo imaging studies employed multi-gradient-echo sequences with various in-plane and through-plane resolutions with minimal SNR constraints (Unconstrained Acquisitions) and with in-vivo time constraints (Constrained Acquisitions), and (iii) in-vivo T2* imaging at standard spatial resolution (1.6×1.6×6 mm³). Detection performance was assessed using receiver operating characteristics (ROC) analysis and its area-under-the-curve (AUC), along with SNR and CNR. Histopathology and mass spectrometry validated iron distribution.

Results:

Simulations demonstrated that finer voxels minimized partial volume but reduced SNR, while coarser voxels increased SNR but masked focal lesions. Optimal performance (AUC=0.80) was identified at 1.3–1.6 mm² pixel size with 2–4 mm slice thickness under Constrained Acquisitions. Detection accuracy declined sharply for iron depositions that were thin in the transmural direction (<1 mm) with limited pixel representation at clinical resolutions. Mass spectrometry (ICP-MS) confirmed significantly elevated iron content in hMI regions vs. remote myocardium (45.1 ± 9.3 vs. 35.3 ± 4.7 μg/g, p=0.012) and correlated with the transmural width of hMI remnants (R² = 0.83).

Conclusion:

Spatial resolution critically determines the detectability of chronic iron remnants on T2* maps. While current clinical protocols can identify larger deposits, small (<1 mm) yet proinflammatory lesions may be missed. These results highlight the need for technical improvements to enhance detection.

INTRODUCTION

Intramyocardial hemorrhage (IMH) is a severe form of reperfusion injury in acute myocardial infarction (MI). It has been identified as the strongest predictor of adverse outcomes in patients with reperfused myocardial infarction (MI) [1–4]. IMH is associated with larger infarcts [5–8], adverse left ventricle (LV) remodeling [9–12], arrhythmias [13–15], poor prognosis [16,17], and persistent iron deposition. The iron deposits have been shown to delay infarct healing and promote adverse remodeling [18–21]. Importantly, even highly localized iron deposits (0.01 mm²) have been shown to trigger disproportionately large inflammatory responses in the infarcted myocardium [19]. Thus, accurate detection of iron deposits within the infarcted myocardium may be essential for accurate diagnosis, as well as development and management of novel therapies.

T2* cardiac MRI is the non-invasive standard for assessing iron remnants in hemorrhagic myocardial infarction (hMI) [13,22,23]. However, the limits of detection of these iron remnants with the current protocols is not known. In chronic MI, iron can appear as a thin band in the transmural direction (<1 mm), especially when compounded by infarct thinning [24]. Further, while several studies have embraced the importance of spatial resolution for detecting iron outside the heart [25,26], most studies investigating iron deposition in hMI have not appraised their findings in relation to imaging resolution (Table 1): in-plane resolutions typically range from approximately 1.1–2.6 mm, with through-plane resolutions of 5–8 mm, resulting in voxel sizes spanning more than a four-fold range. Spatial resolution of cardiac MRI can directly impact both the sensitivity and specificity for detecting focal lesions. Specifically, finer resolution reduces partial volume effects but lowers signal-to-noise ratio (SNR), but coarser resolution improves SNR at the cost of missing focal lesions. This is captured by the well-known relation between SNR and spatial resolution:

$$SNR\propto \frac{\sqrt{N_{acq}}\cdot \sqrt{{\mathrm{V}}_{\text{voxel}}}}{\sqrt{B{W}_{\mathit{\text{read}}}}},\text{where}{\mathrm{V}}_{\text{voxel}}=\mathrm{\Delta }x\mathrm{\Delta }y\mathrm{\Delta }z$$ (1)

where $\mathrm{\Delta }x$, $\mathrm{\Delta }y$, and $\mathrm{\Delta }z$ are voxel size in frequency encoding, in-plane phase encoding and slice-encoding directions, respectively; $N_{\mathit{\text{acq}}}$ represents number of acquisitions (averages); and $B{W}_{\mathit{\text{read}}}$ represents readout bandwidth. Thus, it is not surprising that balancing of spatial resolution and SNR is crucial for optimal T2* imaging of iron remnants. However, whether this is a limitation in imaging chronic iron remnants from hMI with the standard T2* imaging protocols is unclear. Should the current imaging resolutions prove insufficient, it could result in substantial underestimation of focal iron deposits in chronic hMI. In clinical investigations, this potential limitation can translate into underestimation of major adverse cardiovascular events (MACE) risk associated with iron deposition and complicate the assessment of therapeutic efficacy of evolving therapies targeting hMI [27].

Table 1.

Summary of Spatial Resolution Used in Previous T2*-based Studies in detecting hemorrhagic remnants in chronic myocardial infarction.

Manuscript	Species	Spatial resolution			Time after reperfusion
		In-plane	Through-plane	Voxel size
Ghugre et al. [49]	Swine	1.9x1.9 mm2	5 mm	18.05 mm3	6 weeks
Cokic et al. [15]	Canine	1.3x1.3 mm2	6 mm	10.14 mm3	16 weeks
Kali et al. [18]	Canine	1.4x1.4 mm2	6 mm	11.76 mm3	8 weeks
Guan et al. [50]	Canine	1.1x1.1 mm2	6 mm	7.26 mm3	6 months
Bulluck et al. [20]	Human	Not reported			5-7 months
Carberry et al. [21]	Human	2.6x1.6 mm2	8 mm	33.28 mm3	6 months
Carrick et al. [51]	Human	2.6x1.6 mm2	8 mm	33.28 mm3	7 months
Mayr et al. [52]	Human	1.6x1.6 mm2	8 mm	20.48 mm3	9 years
Reinstadler et al. [1]	Human	1.6x1.6 mm2	8 mm	20.48 mm3	12 months
Guan et al. [50]	Human	1.6x1.6 mm2	8 mm	20.48 mm3	6 months
Kali et al. [19]	Human	1.3x1.3 mm2	8 mm	13.52 mm3	6 months

We systematically investigated the influence of spatial resolution on the detection and characterization of residual iron in chronic MI using animal models. Based on the contradictory data in the literature, we hypothesized that the accuracy of detecting iron remnants from chronic hMI can be compromised by strong partial volume effects or low SNR at the standard spatial resolutions often employed in T2* MRI. We investigated our hypothesis using digital phantom simulations and experimental studies in a swine model of hMI.

MATERIALS & METHODS

Numerical Simulations

Left ventricle of the heart was modeled as a truncated 3D ellipsoid [28] with the subendocardial region enriched with iron (width ≈ 1.0 mm; transmurality = 2.5%) to model the iron distribution in chronic hMI. Following imaging parameters were used [29,30]: Infarct zone with T2* = 9.4 ms and T1 = 800 ms; remote (non-infarcted) zone with T2* = 40.1 ms, T1 = 1125 ms). Bloch equation simulations of spoiled gradient-recalled imaging were performed with five echoes (see Table 2 Constrained Ex-vivo Group) and multiple Gaussian noise levels were added (See Supplementary Material 1 for details). T2* maps were fitted with a most widely used 2-parameter mono-exponential model [31], consistent with standard cardiac T2* mapping employed in practice. And three SNR levels were retrospective selected as low (SNR from 1.2 ± 0.0 to 9.1 ± 1.0), intermediate (2.3 ± 0.1 to 22.6 ± 3.2) and high (31.4 ± 0.4 to 317.5 ± 7.6) according to the International Organization for Standardization (ISO) standards (acceptable quality: 10; Excellent quality: 40) [32]. To study partial volume and SNR trade-offs, simulated images were averaged to voxel sizes spanning 0.4, 0.6, 0.8, 1.0, 1.2, 1.6 to 2.0 mm in-plane (isotropic) and 2, 4, 6, 8 mm through-plane at basal, mid-ventricular, and apical locations. SNR of remote myocardium was computed and was used to determine the detectability of iron with Receiver operating characteristic (ROC) and Area-Under-the-Curve (AUC) across all voxel sizes and SNR levels. Refer to Supplementary Material 1 for additional details.

Table 2. Imaging Parameters.

A) Ex-vivo scan parameters for Unconstrained Acquisitions, Constrained Acquisitions and In-vivo Acquisition, respectively. B) In-plane resolution of Unconstrained was 0.3×0.3 mm² - 2.1×2.1 mm², Constrained was 0.8×0.8 – 2.1×2.1 mm² and In-vivo acquisitions was 1.6×1.6 mm².

A
	Unconstrained Acquisition	Constrained Acquisition	In-vivo
TR (ms)	788	80	83.51
TEs (ms)	4.95, 13.70, 22.45, 31.20, 39.95	2.55, 5.80, 9.90, 15.56, 21.22	1.41, 2.64, 3.87, 5.10, 6.33, 7.56,
Flip Angle (°)	60	18	12
Bandwidth	250	780	1185
Average number	16	1
iPAT factor	NA	2
Partial Fourier	NA	6/8
ECG-gated	No	Yes
Slice thickness	2, 4, 6, 8		6
Gradient polarity	Bipolar

B
	In-plane Resolution (mm2)
Unconstrained Ex-vivo	0.3×0.3	0.6×0.6	0.8×0.8	1.0×1.0	1.3×1.3	1.6×1.6	2.1×2.1
Constrained Ex-vivo	NA		0.8×0.8	1.0×1.0	1.3×1.3	1.6×1.6	2.1×2.1
In-vivo						1.6×1.6

Animal Protocol and Image Acquisition

The animal study protocol was approved by the Institutional Animal Care and Use Committee (IACUC, protocol number: 5257). Domestic swine (n = 10, 20-25 kg at baseline) were sedated, intubated and anesthetized and subsequently underwent complete occlusion of the left anterior descending coronary artery (LAD) with a balloon catheter and reperfused after 90 min. Animals underwent cardiac MRI at 8 weeks post infarction and then euthanized and the hearts were excised, rinsed, and preserved in formalin (Sigma-Aldrich, MO, USA) and imaged as described below.

All cardiac MRI was performed on a 3T whole-body MRI system (Siemens MAGNETOM Verio, Erlangen, Germany). For ex-vivo imaging, a representative short-axis slice was identified and imaged with 2D multi-gradient-echo (mGRE) sequence, with varying in-plane and through-plane spatial resolutions (for scan parameters see Table 2) using a 20-channel head coil. Two ex-vivo imaging strategies were employed: (1) Constrained Ex-vivo Imaging designed to imitate conventional in-vivo cardiac MRI acquisitions, and (2) Unconstrained Ex-vivo Imaging designed to isolate the effects of partial-volume averaging independent of SNR limitations. Constrained Ex-vivo Imaging was performed using conventional bright-blood ECG-gated mGRE acquisition parameters, including segmented acquisition and parallel imaging to emulate practical in-vivo cardiac MRI conditions with clinically feasible scan durations. A simulated electrocardiogram (ECG) signal at 60 beats-per-minute was used, with in-plane resolutions ranging of 0.8 × 0.8 mm², 1.0 × 1.0 mm², 1.3 × 1.3 mm², 1.6 ×1.6 mm² to 2.1× 2.1 mm² with the same in-plane field-of-view (FOV) (see Table 2B) and through-plane resolutions of 2, 4, 6, 8 mm (see Table 2A). The Constrained Ex-vivo Imaging were designed to maintain acquisition time within 20 seconds per acquisition, feasible for practical breath-hold cardiac MRI conditions. Unconstrained Ex-vivo Imaging was performed without ECG gating or acceleration constraints and with 16 signal averages to maximize SNR and spatial resolution. These acquisitions were not intended to represent clinically practical in-vivo imaging conditions, but rather to provide a high-SNR/spatial resolution reference framework for isolating the effects of partial volume from spatial resolution and SNR in detecting iron deposits. Images were acquired with high averages (16 averages) without ECG gating at in-plane resolution of 0.3×0.3 (hardware limitation) to 2.1×2.1 mm² with fixed FOV (see Table 2B) and through-plane resolution of 2 to 8 mm (in 2-mm increments), see Table 2A. In both Constrained and Unconstrained acquisitions, the TEs and TRs were minimized to its hardware limitation and RF excitations were performed at the Ernst angle [33] to maximize signal. In-vivo Imaging: 10 animals underwent breath-held (~10 seconds), ECG-gated, short-axis, T2* cardiac MRI at end diastole with 18-channel surface coil (See Table 2A). The details of the Constrained Ex-vivo, Unconstrained Ex-vivo and In-vivo acquisition time for per each in-plane resolution are shown in Table 3.

Table 3.

Acquisition Time for Unconstrained Acquisitions, Constrained Acquisitions and In-vivo Acquisition.

In-plane Resolution (mm2)	0.3×0.3	0.6×0.6	0.8×0.8	1.0×1.0	1.3×1.3	1.6×1.6	2.1×2.1
Unconstrained Ex-vivo (min)	120.1	55.5	40.4	30.3	25.2	20.2	15.2
Constrained Ex-vivo (sec)	NA		17	14	12	10	9
In-vivo (sec)						10

Histopathological Validation and Mass Spectrometry

Following imaging studies, hearts were sliced into 2-mm thick short-axis rings and were matched to MR images based on fiducial landmarks (right ventricular insertion points, papillary muscle positions, etc.). One slice was used for histopathology analysis, and the other was used for mass spectrometric (ICP-MS) analysis. For histopathology, a region that includes the whole MI zone and remote myocardium were sectioned and stained with hematoxylin and eosin (H&E, for tissue damage), Masson Trichrome (MT, for fibrosis), and Prussian Blue (for iron deposits). The stained sections were digitally imaged at 10- and 40-fold magnifications. For ICP-MS analysis, corresponding slices were further sectioned resulting in 2 remote and 2 MI sections. Iron content within each of these sections was determined using standard ICP-MS (NexION 2000, PerkinElmer) [18].

Cardiac MRI - Image Analysis

Image analysis pipeline is provided in Fig. 1. All image analysis was performed in MATLAB 2021a (MathWorks, Natick, Massachusetts). Endo- and epi-cardial contours were used to delineate the left-ventricular myocardium. Reference ROIs (MI, remote) were drawn on 0.3×0.3×2 mm³ T2* maps by two experienced cardiac MRI readers in consensus (Unconstrained ex-vivo images), scaled and copied onto corresponding images at alternate resolutions. T2* maps were fitted using qMRLab (http://qmrlab.org/) and goodness-of-fit (R²) and SNR of remote myocardium were calculated to assess quality of T2* maps based on:

$$SNR=\frac{\mu \left(\mathit{\text{remote}}\right)}{\sigma \left(\mathit{\text{remote}}\right)},$$ (2)

Where μ(remote) and σ(remote) denote mean and standard deviation (SD) of the signal intensity of the remote myocardium on T2* maps.

Figure 1. Image Processing Pipeline to Determine the Capability of Ex-vivo and In-vivo T2*.

The T2* map acquired at a spatial resolution 0.3×0.3×2 mm³ with 16 averages (Unconstrained Acquisition) was chosen as reference standard. Contours (epicardium and endocardium, MI, and remote myocardium) were drawn and binary masks were created. The myocardium was evenly divided into 50 segments and applied to MI mask. Iron deposits were identified within MI region using Mean-2SD method, converted to a segmental percentage of iron. The mean iron transmurality and the width of the iron deposition were also computed. Finally, the array of percentage of iron was converted to binary array at 1% cutoff as reference standard in ROC analysis. For the remaining ex-vivo images, contours drawn from the reference standard were scaled accordingly and binary ROI masks were generated. As with reference images, each image was divided evenly into 50 segments for analysis and analyzed similarly. For the In-vivo images underwent similar analysis. Percentage of iron volume for both in-vivo and reference standard (RS) was computed as V(Invivo) and V(RS) respectively. The bias of the percentage of the iron volume between in-vivo and ex-vivo ground truth was computed to assess the accuracy of iron in-vivo.

Detection Accuracy:

For ex-vivo analyses, the 0.3 × 0.3 × 2 mm³ dataset from Unconstrained ex-vivo images served as the reference standard. For regional analysis, the myocardium was subdivided into 50 circumferential segments per slice to provide higher spatial resolution compared to the conventional AHA 17-segment model. This approach is consistent with prior transmurality analyses and was considered appropriate for the ex-vivo setting, where the heart was imaged in a fixed position without motion, enabling more detailed spatial characterization of iron distribution [34]. Iron-rich MI zones were delineated using mean-2SD criterion [35,36] and the percentage of MI zone with iron calculated for each segment, and then converted to binary arrays as iron-positive or iron-negative zones with a threshold of 1%. A 1% cutoff was used to eliminate segments with spurious hyperintense pixels from being considered as iron-positive. Patient-level bootstrap ROC analysis (2000 sampling points) was used to obtain robust estimates of performance of AUC to account for the hierarchical structure of the data, based on iron mask of reference standard as determined from Unconstrained ex-vivo images acquired at 0.3 × 0.3 × 2 mm³ voxel size. In addition, the width of the iron-rich region (mean transmurality and transmural width) was also computed.

In-vivo vs. Ex-vivo:

Volume fraction of iron based on T2* maps within the slice of interest determined from in-vivo and ex-vivo images and then compared. Using Unconstrained ex-vivo images (at 0.3 × 0.3 × 2 mm³) as the ground truth, the bias of the percentage of the iron volume were computed (Fig. 1).

Image Contrast:

To assess the interplay between SNR and partial volume effects in iron detection, binary iron-rich MI zones were determined from T2* maps acquired at 0.3 × 0.3 × 2 mm³ from Unconstrained acquisitions (reference standard) with Mean-2SD criterion [18]. The iron-rich zones and previously drawn remote regions of interest (ROIs) were downsized to match each spatial resolution of T2* maps from the remaining Unconstrained and all Constrained T2* maps. Binary masks were downsampled using bicubic interpolation to minimize aliasing and preserve anatomical boundaries during resolution reduction. The resulting probabilistic masks were re-binarized using a threshold of 0.25, corresponding to a partial-volume criterion that retains voxels containing ≥25% of the original high-resolution mask. The downsized iron masks served as the ground truth of iron-rich regions. Contrast-to-noise-ratio (CNR) of remote myocardium to iron remnants was then computed for both Constrained and Unconstrained T2* maps:

$$CNR=\frac{\mu \left(\mathit{\text{remote}}\right)-\mu \left(\mathit{\text{downsized iron mask within MI}}\right)}{\sigma \left(\mathit{\text{remote}}\right)},$$ (3)

where μ(remote) is the mean and σ(remote) is the standard deviation of remote myocardium while μ(downsized iron mask within MI) downsized iron mask of T2* maps from 0.3×0.3×2 mm³ reference standard.

Statistical Analysis

All statistical analysis was performed using MATLAB 2021a (MathWorks, Natick, Massachusetts) and Prism 9.2.0 (GraphPad, San Diego, California). The normality test of continuous variables was performed using Shapiro-Wilk test. Normally distributed continuous variables are reported as mean ± SD, while non-normally distributed variables (AUC) reported as mean with 95% confidence interval (CI). The differences in iron concentration between groups were evaluated using Student’s t-test. ROC analysis of iron remnant detection on 50 segments was performed per subject, and the corresponding AUC were computed for linear regression analysis between width of iron deposits vs AUC of ROC curves. Association between width of iron deposits vs. iron concentration from ICP-MS was also assessed using Pearson correlation coefficients and compared using linear regression. Repeated-measure two-way ANOVA was used to assess AUCs determined for in-plane and through-plane resolutions. F-test was conducted to compare variances of detection bias between sub-groups. Significance was established for P < 0.05 and indicated by *. P < 0.01 and P < 0.001 are denoted as ** and ***, respectively. Absence of statistical significance is denoted as (NS).

RESULTS

Numerical Simulations – SNR vs Spatial Resolution:

We simulated T2* maps with in-plane resolution of 0.32 (0.4×0.4×2) mm³ to 32 (2.0×2.0×8) mm³ at 3 representative SNR levels to illustrate the effect of spatial resolution and SNR: low SNR (ranging from 1.2 ± 0.0 to 9.1 ± 1.0), intermediate SNR (ranging from 2.3 ± 0.1 to 22.6 ± 3.2), and high SNR (31.4 ± 0.4 to 317.5 ± 76.0). Representative short-axis T2* maps (basal, mid-ventricle, and apical) with spatial resolutions of 0.4×0.4×2 mm³ and 2.0×2.0×8 mm³ at the three SNR levels with iron deposits are shown in Fig. 2. The findings here support the notion that the visibility of iron increases with SNR at the expense of partial volume effects with the increase in voxel size. This is further confirmed by the AUC measures from ROC analysis (Fig. 2). Specifically, at the high SNR level, AUC decreased with voxel size in the in-plane dimension. A similar trend was also observed in the through-plane dimension, particularly at the apex. These findings identify that partial volume plays a key role in the detection of chronic iron remnants in hMI. In addition, our findings show that at each slice location, as the noise level decreases, AUC increases. Notably, at the intermediate SNR and moderate voxel sizes (≈1–10 mm³), optimal AUCs were found suggesting a balance between SNR and partial volume effects. Our findings here lay the foundation for the notion that the compromise between partial volume effect and SNR limits associated with spatial resolution are critical for maximizing detectability of chronic iron deposits in hMI.

Figure 2. Numerical Phantoms - AUC from ROC Analysis Targeting Dectection of Thin Iron Bands at Base, Mid-ventricle and Apex at Three Different SNR Levels.

Left column shown the schematic of the base mid-ventricle and apex with thin iron bands. Nine scatter plots are shown: horizontal direction shows data from low to high SNR and vertical direction shows data from base to apex. Each scatter plot shows voxel size vs. AUC from the respective ROC curves, where through-plane thicknesses of 2, 4, 6, 8 mm are denoted as circle, square, diamond and triangle, respectively. Optimal voxel size range is identified using 95% of maximum AUC (AUCmax) as cutoff (shown by dotted horizontal red lines). The cutoff values are displayed in each plot. Representative T2* maps are shown at 0.4 × 0.4 × 2 mm³ and 2.0 × 2.0 × 8 mm³ resolutions at low, intermediate and high SNR levels from the base, mid-ventricle, and apex. The data from above demonstrate the influence of spatial resolution on SNR and partial volume effect in the detection of iron-rich band.

Animal Studies: Ex-vivo Findings

(i). Transmurality and Width of Hemorrhagic Remnants in Chronic MI:

Iron deposition at 8-weeks post MI appeared as narrow subendocardial bands (Supple. Fig. 1). Representative T2* maps (one with a narrow band of iron deposition and the other with significantly broader band of iron deposition) are shown in Supplementary Fig. S1A. Note that at in-plane resolutions between 0.8–2.1 mm², this corresponds to 1–4 pixels across the iron band supporting the notion that spatial resolution of conventional T2* imaging can be significantly affected by partial volume effects.

(ii). Partial Volume Effect:

To isolate the partial volume effect, we employed Unconstrained Acquisitions, which allowed us to investigate changes in spatial resolution while ensuring adequate SNR. Representative images from two pigs (swine 1 with narrow distribution of iron (Supplementary Fig. S2A); swine 2 with broader distribution of iron (Supplementary Fig. S2B)) show that the conspicuity of iron decreases with increasing voxel size by visual inspection, independent of iron distribution. Aggregate data were consistent with representative findings with declining AUC with increasing voxel size in both in-plane and through-plane dimensions (Supplementary Fig. S2C). Notably, a steeper decrease in AUC was found with decreasing in-plane resolution. At 2-mm through-plane resolution, AUC decreased from 1.00 (95% CI: [1.00, 1.00]) to 0.70 (95% CI: [0.61, 0.78]) with decrease in in-plane resolution from 0.3×0.3 mm² to 2.1×2.1 mm². While at 8-mm through-plane resolution, AUC decreased from 0.74 (95% CI: [0.66, 0.85]) to 0.66 (95% CI: [0.58, 0.75]) with decrease in in-plane resolution from 0.3×0.3 mm² to 2.1×2.1 mm² (See Supplementary Fig. S2D), indicating that partial volume plays a significant role in both in-plane and through-plane resolutions.

(iii). In-Plane vs. Through-Plane Resolution:

To further investigate the importance of spatial resolution in detecting chronic iron deposits, we performed Constrained Acquisitions. Representative images from two swine (one with narrow distribution of iron (Fig. 3A); the other with broader distribution of iron (Fig. 3B)) showed a decrease in visual conspicuity of iron-rich regions on T2* maps with increasing voxel size. However, the noise level of T2* maps decreased as voxel size increased, which also affects the detection of iron-rich regions (1.6×1.6×2 vs. 1.6×1.6×6 mm³ in Fig. 3A and B). Figure 3C shows that both through-plane and in-plane resolutions significantly affected detection, but the effects were more pronounced with changes in through-plane resolution than with in-plane resolution (from Two-way ANOVA). Optimal detection (AUC of ROC at 95% cutoff) occurred at 1.6×1.6×2 mm³ (AUC = 0.82, 95% CI: [0.72, 0.88]) and 1.3×1.3×4 mm³ (AUC = 0.80, 95% CI: [0.70, 0.86]) (Fig. 3D). This result is also supported by the simulation phantom at intermediate SNR level, which showed that detection of iron remnants can be optimized through optimization of spatial resolution to balance SNR and partial volume effects (Fig. 2).

Figure 3. Constrained Acquisition - AUC from ROC Analysis Targeting Dectection of Thin Iron Bands.

Representative examples A) Swine 1 and B) Swine 2. Reference T2* maps acquired with spatial resolution of 0.3×0.3×2 mm³ (Unconstrained Acquisition) and T2* maps acquired with Constrained Acquisitions are shown at three different spatial resolutions: Left to right are 1.6×1.6×2, 1.6×1.6×6 and 2.1×2.1×8 mm³. Iron-rich regions are identified with arrows heads on reference images. These iron-rich regions were identified by Mean-2SD method and highlighted on the row of “Processed” as magna masks in magnified view of T2* maps focusing MI region (Blue dotted box from raw image). C) Corresponding ROC curves and the associated AUC at each spatial resolution is shown for Constrained group with the reference T2* maps serving as the ground truth. Two-way ANOVA analysis of in-plane resolution and through-plane resolutions are shown in the immediate table below. AUC is shown along with 95% Confidence Interval (CI) for each of the ROC curve with its background color coded from 0.5 to 1.0. D) Range of optimal voxel size range permitting the identification of iron from Constrained Acquisition is shown by a loop encompassing the data points at spatial resolutions of 1.6×1.6×2 and 1.3×1.3×4 mm³ with 95% AUC as a cutoff (0.8 – 0.82). AUC corresponding to various through-plane resolutions (2, 4, 6, 8 mm) are identified as different shapes (circle, square, diamond and triangle, respectively). E) Scatterplot showing the relationship between mean AUC (derived from Constrained Acquisitions) and width of iron deposition is shown. AUC was directly related to width of iron deposition (Y = 0.50X + 0.16, R²=0.84, P<0.001), with AUC for data acquired with Swine 1 is 0.56 (width of iron deposition = 0.59 mm) while AUC for Swine 2 is 0.85 (width of iron deposition = 2.40 mm).

(iv). Spatial Extent of Iron Distribution:

We found that hMIs with thin (less transmural) iron regions can be challenging to detect with T2* (See Fig. 3E, R² = 0.84, P < 0.001). When the animals were equally divided into 2 groups based on the width of iron deposition (group 1:1.27 ± 0.66 mm vs. group 2: 0.57 ± 0.03 mm respectively, P = 0.04), smaller AUC of detection was associated with iron deposits with narrower bands than with wider bands (0.77 ± 0.09 vs. 0.90 ± 0.09, P = 0.02).

(v). CNR Analysis:

Representative examples of an animal with a thin iron band (Swine 1; 0.59 mm) and broader iron band (Swine 2; 2.4 mm) are shown in Supplementary Fig. S3A and S3B respectively, at high and low spatial resolutions. These representative cases shows that that CNR is strongly dependent on the extent of the iron band. The CNR computed from the downsized iron masks showed negative image contrast for thin bands, indicating strong partial volume effects (Supplementary Fig. S3A). From these analyses, we identified subjects with CNR > 2 (robust; Supplementary Fig. S3C) and CNR > 1 (acceptable; Supplementary Fig. S3D). The highest percentage for acceptable contrast (CNR > 1) occurred at 5.12 mm³ (1.6 × 1.6 × 2 mm³): 60%; and highest percentage for robust contrast (CNR > 2) occurred at 6.76 mm³ (1.3 × 1.3 × 4 mm³): 30% subjects. These results align with optimal iron detection from Fig. 3D and highlights that even with optimal achievable spatial resolution, the ability to detect iron bands in severely thinned hMI can be limited.

(vi). Histology and ICP-MS Validation:

Histopathology confirmed iron deposition in chronic MI in all subjects (n=10). Histological findings from the two representative subjects (same subjects as in Fig. 3) with chronic MI are shown in Fig. 4 (Swine 1: panels A-C; and Swine 2: D-F), where contiguous H&E sections, Prussian Blue sections, and MT-stained sections show colocalization of tissue damage (necrosis), iron deposition and fibrotic region respectively. MI territories were stained as both necrotic and fibrotic tissue (Fig. 4B and Fig. 4E), with scattered iron deposition within the MI region. More iron-stained regions were seen in subjects with wider iron bands (larger in swine 2 than in swine 1, Fig. 4C and Fig. 4F), which is consistent with our MRI findings. ICP-MS showed significantly higher iron in MI regions vs. remote regions (n = 10, 45.1 ± 9.3 vs. 35.3 ± 4.7 μg/g, P = 0.012, Fig. 4G), confirming the presence of iron deposition in the MI zone. Iron concentration correlated strongly with width of the iron band (R² = 0.83, P < 0.001, Fig. 4H), indicating a strong association between iron concentration (i.e. 1/T2*) and the extent of iron deposition.

Figure 4. Representative Histology and ICP-MS from Swine Hearts.

Histological findings from swine 1 and seine 2 (from Figs. 3 and 4) are shown: A) and D) are H&E sections; and B) and E) are Trichrome stained sections. These sections show thinning of the infarct sections consistent with MRI findings (See Fig. 3) that are site of replacement fibrosis. Prussian blue staining of the heart from Swine 1 C) shows relatively low abundance of iron inside the infarct, while Swine 2 F) shows significant iron accumulation. Small rectangular arrows pointing to remote and MI zones in the 10× view images (A, B and C) are magnified on the right. Iron depositions are identified with black arrow heads on zoomed-in Prussian blue images. G) Shows bar plot of iron concentration in remote myocardium vs. MI zone. * denotes P < 0.05. H) Linear regression of width of iron deposition vs the ratio of iron concentration in MI ([Fe]_MI) and iron concentration in remote ([Fe]_Remote), showed a strong correlation (Y = 0.60X + 0.73, R² = 0.83).

Animal Studies: In-vivo Findings

In-vivo cardiac MRI performed with a standard spatial resolution (1.6×1.6×6 mm³) was analyzed and compared with the reference standard from Unconstrained Acquisition at 0.3×0.3×2 mm³. Our in-vivo findings were inconsistent with the extent of the iron observed at the reference resolution of Unconstrained Acquisition. Specifically, the bias was significantly higher for animals with narrower band of iron within MI compared to those with broader band of iron deposition (Fig. 5A and 5B). Three representative examples of over-, correct-, and underestimation of iron rich regions within MI are shown in Fig. 5C. These findings demonstrate the limits of detection of iron rich regions within chronic hMI based on the current cardiac MRI protocol.

Figure 5. Limitations of Standard T2* Cardiac MRI in Detecting Chronic Iron Deposits.

A) Scatter plot of percentage of iron volume bias vs. width of iron deposit (mm) is shown. Consistent with expectations, as the width of the iron deposition is increased bias is observed to decrease. B) Animals divided into two groups, one with low iron and the other with higher iron content (details in text), showed that the bias was significantly lower in the animal group higher iron than low iron. C) Three representative examples from the 10 animals with an over-estimation of iron (Swine 1), low-bias (Swine 2) and under-estimation of iron (Swine 3, where the iron-rich is completely undetected) are shown.

DISCUSSION

We systematically investigated the importance of spatial resolution on T2*-based detection of chronic iron in hMI. Our numerical simulations showed that both in-plane and through-plane spatial resolutions are key determinants for accurate detection of iron deposits. We validated our numerical findings using ex-vivo pig hearts with chronic hMI, enabling evaluation of resolution–SNR trade-offs under controlled, motion-free conditions. Our results suggest that an appropriate balance between spatial resolution, SNR, and partial volume effects is important for detecting iron deposits following hemorrhagic MI, although detectability may also be influenced by image noise and the size and spatial distribution of iron deposits. In addition, our in-vivo study, based on a representative single acquisition protocol (1.6 × 1.6 × 6 mm³), suggests that the detectability of chronic iron deposits may be limited under commonly used imaging conditions. Overall, these findings indicate that the performance of T2* mapping in detecting iron deposits is strongly tied to spatial resolution, noise, and the spatial extent of the iron deposition.

We observed from both simulations and imaging studies that thin iron bands (<1 mm) were poorly detected even at high spatial resolution, while broader iron bands yielded higher AUC and CNR. Iron concentration measured by ICP-MS correlated strongly with the width of the iron bands, underscoring the practical challenge in detecting these thin bands within the chronic MI territories. Our stimulations also provide insight into the relations between detectability and the location of iron deposition or hMI. We found that when hMI is localized to apical regions, the partial volume effect from through-plane resolution is stronger, which can reduce the detection accuracy. Compared to simulations, where we modeled the iron-rich regions as a 1-mm homogeneous layer, in-vivo, it may be more complicated as iron is likely scattered making partial volume effects to be even more pronounced. Therefore, the simplified model used in this study provides a basis for evaluating resolution–SNR trade-offs, with the understanding that it does not fully capture the complexity of in-vivo iron distribution. Thus, the reported detectability limits should be interpreted as reflecting idealized, controlled conditions where the effects of spatial resolution and SNR are isolated from complex physiological and susceptibility-related confounders.

Susceptibility-induced blooming effects may also influence the apparent size of iron deposits in T2* imaging. Local magnetic field perturbations caused by iron deposits can lead to signal loss extending beyond the true physical boundaries of the deposition, resulting in an apparent enlargement of hypointense regions. This effect may partially counteract partial volume limitations by increasing conspicuity, but can also lead to overestimation of lesion size and reduced spatial specificity, particularly at lower spatial resolutions. Therefore, the observed detectability and spatial extent of iron-rich regions reflect a combination of true tissue properties and susceptibility-mediated field shifts that affect gradient-echo signals.

The relationship between SNR of remote myocardium of T2* maps and voxel size across simulations, ex-vivo and in-vivo images provides additional insights (See Supplementary Fig. S4). Notably the slope of the regression between SNR and $\sqrt{\text{voxel size}}$ reflects factors independent of spatial resolution (other imaging parameters, field strength, coil geometry and selection etc.); a steeper slope indicates greater ability to increase SNR solely by adjusting voxel size. For the Constrained acquisitions we found this slope to be 5.24 (R² = 0.935, P < 0.001), while simulations at low and intermediate SNR levels yielded slopes of 4.08 (R² = 0.997, P < 0.001) and 1.43 (R² = 0.990, P < 0.001), respectively. The slightly higher slope between SNR and $\sqrt{\text{voxel size}}$ from the Constrained acquisitions compared favorably with the simulation data determined using intermediate noise level (see Figs. 2 and 3E). For practical reasons only one parameter set was acquired in the in-vivo study (i.e. a full SNR profile vs $\sqrt{\text{voxel size}}$ was not possible), the measured SNR (7.5 ± 2.1 at 1.6 × 1.6 × 6 mm³) was found to be between the simulated low-SNR level (5.6 ± 0.4) and intermediate-SNR level (16.8 ± 1.4), and well below the constrained ex-vivo level (21.8 ± 3.1). The comparatively lower SNR and limited detection of particularly iron bands in the in-vivo scans may be attributable to motion (cardiac/respiratory motion and blood flow), off-resonance effects around the heart, and coil sensitivity. These findings indicate that the current SNR of in-vivo acquisitions are in a range where it will be challenging to detect the thin iron deposits. This also means that strategies which can increase SNR of in-vivo acquisitions can open doors for improving the accuracy for detecting narrower bands of iron depositions.

We found high diagnostic performance (AUC > 0.8) of ex-vivo T2* maps at the optimal spatial resolution of 1.6×1.6×2 mm³ and 1.3×1.3×4 mm³. This aligns well with our simulation at the intermediate SNR level at the apical regions with through-plane resolution of 2 and 4 mm. These findings indicate suggest that imaging along the long axis direction of the heart can provide additional insight, which motivates the need for developing high, isotropic, resolution 3D T2* mapping techniques. Our CNR analysis further validates the connection between the optimal resolution and the contrast between iron-rich region and remote myocardium.

Consistent with previous studies, we used Mean-2SD criterion to identify the iron deposits in T2* maps. This approach provides a subject-specific threshold that accounts for background signal variability and image noise. Alternate approaches, such as those that rely on T2* < 20 ms may lead to different results, however, it may also introduce inter-subject variability [35] and would not take image noise into consideration.

We exploited our 3T MRI scanner with the highest prescribable spatial resolution (0.3×0.3×2 mm³) with high averages (16 averages) to serve as the surrogate reference standard for detecting chronic iron deposits under near ideal imaging conditions. These data demonstrated excellent fitting stability [mean goodness of fit (R²): 0.993 ± 0.001] and high SNR (19.7 ± 2.8),. Both of these measures provided the confidence that the reference spatial resolution of Unconstrained ex-vivo images of 0.3×0.3×2 mm³ and the 2-parameter mono-exponential fitting used here are robust to serve as the ground truth. However, we acknowledge that MRI, even at high resolution and high SNR, cannot serve as a true gold standard given the influence of susceptibility effects, partial volume averaging etc. Accordingly, the ex-vivo dataset was used to assess relative detectability across different spatial resolutions and SNR regimes rather than with respect to absolute diagnostic accuracy. Although histology and ICP-MS confirmed the presence of iron deposition in infarcted regions, they were not used for voxel-wise reference masks due to challenges in precise co-registration (e.g., tissue deformation and resolution mismatch), and thus served as qualitative validation. This limitation and the potential for propagation of bias from the reference images is acknowledged.

Enhancing the accuracy for detecting narrow iron deposits in-vivo is currently limited by SNR. This may be addressed in one of several ways. First, current standard T2* mapping requires breath-holding which limits data acquisition. This may be overcome with free-breathing, high resolution, T2* mapping approaches, many of which have been reported [37–40] but their utility for current application has not been investigated. Second, hardware improvements (such as receive coils with increased number of RF channels) [41], optimized receiving gain parameters [42], increased field strength may help to overcome the SNR challenge. Third, improvements in T2* fitting using a more sophisticated fitting strategy, e.g. offset exponential or bi-exponential [43], discarding long echoes where the SNR falls below 2.5 [44] may also prove beneficial. Finally, with the rapid adoption of machine learning in MRI, deep learning based techniques that enhance T2* mapping may prove useful in enhancing detecting of iron deposits following hMI [45]. More broadly, recent MRI-based analysis studies, including radiomics and machine learning approaches, have demonstrated the potential of quantitative imaging features for characterizing tissue properties and predicting clinical outcomes [46,47]. In this context, the present work focuses on foundational basis for understanding how spatial resolution and SNR influence the detectability of susceptibility-based biomarkers such as iron deposition.

Study Limitations

In spite the rigor of this study, it is not without limitations. First, all ex-vivo scans were performed in formalin-fixed hearts, which may affect the T2* values of the myocardium but past studies have shown that even 50 days of formalin fixation does not alter T2* values of hearts [48]. Hence, we expect the T2* values reported in our study to be applicable to all in-vivo scans. Second, we focused on spatial resolution and partial volume effect in detecting iron deposition but not on other potential confounders, such as B₀ off-resonance, cardiac and respiratory motion, susceptibility confounders near the lung, blood flow effects etc. These effects can introduce additional signal variability and spatial heterogeneity, potentially reducing detection sensitivity of iron deposits in clinical settings. Hence even the proposed optimal parameters ex-vivo may not be sufficient to overcome other confounders (off-resonance/motion). Further, the reported detectability limits likely represent an upper bound and may not fully reflect in-vivo performance. Third, our animal studies were limited to left-anterior wall infarctions. Further studies may be needed to confirm whether our conclusions are applicable to other MI territories. Fourth, because our ex-vivo imaging was performed with the heart in a fixed position, spatial misalignment across datasets was expected to be minimal. Nevertheless, transferring ROIs from the highest-resolution dataset to lower-resolution images may introduce minor boundary inaccuracies due to mask resampling, particularly near small iron-rich regions, and this should be considered when interpreting detection performance. This study includes a limited number of animal experiments (N=10), which may restrict the generalizability of the findings. Although consistent trends were observed across the evaluated conditions, larger studies will be needed to further validate these results and assess variability across a broader range of subjects and imaging conditions.

CONCLUSION

We conclude that the spatial resolution of T2* maps is a critical factor in the detectability of iron remnants following chronic hemorrhagic MI. Under commonly used T2* MRI conditions, large-scale iron deposits can be identified, whereas the detectability of small deposits (< 1 mm³, but also highly proinflammatory) may be limited. These findings highlight the influence of spatial resolution, noise, and lesion size on T2* mapping performance, and suggest that technical improvements may further enhance the detectability of chronic iron deposits based on T2* MRI.

ABBREVIATIONS

AUC

Area under the curve

CMR

Cardiovascular magnetic resonance

CNR

Contrast-to-noise ratio

ECG

Electrocardiogram

EMT

Elastin Masson’s Trichrome

H&E

Hematoxylin and Eosin

ICP-MS

Inductively coupled plasma mass spectrometry

IMH

Intramyocardial hemorrhage

LAD

Left Anterior Descending

LV

Left ventricle

mGRE

Multi-gradient echo

MI

Myocardial infarction

ROC

Receiver operating characteristic

ROI

Region of interest

SNR

Signal-to-noise ratio

Availability of data and materials

The data that support the findings of the current study are openly available at the following URL/DOI: https://doi.org/10.5281/zenodo.18228221