Automated scar quantification on late gadolinium enhancement cardiac MR in patients with acute myocardial infarction: a cross-modal transfer learning and expert-refined approach

Introduction

Background

Myocardial infarction (MI) remains one of the leading causes of cardiovascular mortality worldwide (1). Late gadolinium enhancement (LGE) cardiac magnetic resonance (CMR) is the clinical reference standard for visualizing and quantifying myocardial scar and fibrosis (2). Accurate assessment of scar burden (e.g., scar size and transmurality) is clinically important for risk stratification, guiding revascularization strategies, and prognosis evaluation (3).

Despite its diagnostic value, quantitative LGE analysis in routine practice still relies heavily on manual delineations. These procedures are time-consuming and prone to substantial inter-observer variability, limiting scalability and longitudinal reproducibility (4). A key prerequisite for automated scar analysis is reliable segmentation of the left ventricular (LV) myocardium on LGE images. However, unlike cine CMR where large annotated benchmarks such as automated cardiac diagnosis challenge (ACDC) dataset exist (5), pixel-wise annotated LGE datasets remain scarce due to low myocardium-blood contrast and frequent artifacts, making it difficult to train robust supervised segmentation models.

Rationale and knowledge gap

Two practical challenges prevent LGE scar quantification from being fully automated and reproducible across centers.

First, annotation scarcity and domain shift limit myocardium segmentation performance in LGE. While deep learning has achieved strong results in cardiac MR segmentation, most approaches depend on abundant labeled data and show degraded performance when transferred across modalities (6). The domain shift between cine and LGE (contrast, noise, resolution, and artifacts) prevents direct reuse of cine-trained models. Unsupervised domain adaptation methods have been explored, but in the absence of explicit supervision or expert guidance they may not reach the boundary accuracy needed for clinically reliable downstream scar quantification (7). Recently, semi-supervised learning methods have been introduced to bridge this gap by leveraging a small amount of labeled LGE data alongside unlabeled data through techniques like pseudo-labeling and consistency regularization (8). However, existing semi-supervised approaches in cardiac magnetic resonance imaging (MRI) often rely heavily on the quality of pseudo-labels, which are prone to severe errors due to low tissue-to-blood contrast, severe class imbalance, and LGE-specific artifacts (9). Furthermore, these methods typically still require a considerable initial annotation effort to achieve robust and stable performance across varying clinical centers, which limits their scalability.

Second, even with accurate myocardial masks, operator-dependent reference region selection remains a major source of variability for threshold-based scar quantification. In LGE, absolute intensities are not standardized and depend on acquisition timing, contrast kinetics, and scanner parameters (10). Therefore, conventional methods such as the n-standard deviation (n-SD) technique and the full-width at half-maximum (FWHM) approach require manual definition of reference regions (e.g., remote myocardium, blood pool, or scar seed) (11). Manual placement of these regions of interest (ROIs) is known to introduce large inter-observer differences and can lead to substantial discrepancies in infarct size estimation (4,12,13), especially when ROIs inadvertently include microvascular obstruction (MVO) or artifact-prone areas.

These gaps motivate a workflow that (I) leverages cine annotations to reduce the LGE labeling burden while maintaining clinically adequate boundary accuracy, and (II) removes subjective ROI placement from FWHM-based scar quantification.

Objective

The objective of this study is to develop and validate a fully automated pipeline for myocardial scar quantification from LGE. Specifically, we aim to:

Obtain accurate LV myocardial masks on LGE by cross-modal semi-supervised learning that transfers anatomical knowledge from cine CMR to LGE via disentangled representations (7,14,15), followed by expert correction and fine-tuning.

Perform operator-independent FWHM-based scar quantification by using a segment-wise enhancement probability predictor to automatically identify remote myocardium and scar seed regions.

Methods

Datasets

Multicenter patient cohorts

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Ren Ji Hospital, Shanghai Jiao Tong University School of Medicine (No. [2018]093), and informed consent was waived due to the retrospective nature of this study. All participating institutions were informed of and agreed to this study. Cine and LGE image pairs were collected from a total of 420 patients with acute myocardial infarction (AMI) across three tertiary grade-A general hospitals in China (hereafter centers 1, 2, and 3): center 1 (internal, n=390) and centers 2/3 (external, n=12 and 18). This cohort substantially overlaps with that of our previous study (16), which enrolled 430 AMI patients from the same three centers; 420 of the 430 patients had complete, paired cine and LGE sequences and were included in the present analysis. Patients were diagnosed with ST-segment elevation MI and underwent CMR within 7 days after percutaneous coronary intervention. Patients with non-ischemic cardiomyopathies or valvular and congenital heart disease were excluded. Demographics of the enrolled patients are summarized in Table S1, and the cine and LGE imaging parameters of the three centers are detailed in Table S2.

In addition, the publicly available ACDC dataset was used exclusively to pretrain a cine-domain segmentation network [nnU-Net (17)] that subsequently produced pseudo-labels on in-house cine images for cross-modal transfer. No ACDC subject appears in any subsequent training or test set.

Data preparation and quality control

The patient and slice selection process is summarized in Figure 1. Throughout this work, one cine-LGE pair refers to a matched cine and LGE short-axis slice from the same patient at approximately the same anatomical level. All images were preprocessed to a unified matrix of 128×128 pixels with an in-plane spacing of 0.89×0.89 mm². Eligibility filtering required (I) availability of paired bSSFP cine and LGE protocols, (II) aligned inter-protocol imaging axes, and (III) inter-protocol slice distance <4 mm. Pairs were subsequently screened by manual quality control to exclude those with cine artifacts, cine-LGE misalignment, LGE artifacts or poor contrast, and overly apical or basal slice positions. Detailed exclusion counts at each step are reported in Figure 1.

Figure 1 Flowchart of multicenter patient selection for developing and validating the proposed pipeline. Patients from center 1 (internal) were screened by automatic eligibility checks and manual quality control, then randomly partitioned at the patient level into a cross-modal adaptation set and an expert-corrected fine-tuning set, both used for model development. Patients from centers 2 and 3 (external) were reserved exclusively as an independent test set. Center 1, Ren Ji Hospital; center 2, Longyan First Hospital; center 3, Beijing Anzhen Hospital. AMI, acute myocardial infarction; bSSFP, balanced steady-state free precession; LGE, late gadolinium enhancement.

Dataset partitioning

After quality control, 281 patients (1,729 cine-LGE pairs) from center 1 and 30 patients (216 LGE slices) from centers 2 and 3 were retained. The center-1 cohort was partitioned at the patient level into two non-overlapping subsets: a cross-modal adaptation set with 218 patients (1,409 slices), used without expert labels for unsupervised dual attention-guided fuzzy network (DAFNet) (7) training and enhancement probability predictor training; and an expert-corrected fine-tuning set with 63 patients (320 slices), in which DAFNet predictions were manually corrected by a CMR expert (D.A.) and used for supervised fine-tuning of the segmentation. The fine-tuning set was used in its entirety for supervised DAFNet fine-tuning and for threshold selection; no internal training/validation subset was retained, in order to maximize the supervisory signal from the limited expert-corrected labels, and model evaluation was performed exclusively on the held-out external test set (centers 2/3). The external set with 30 patients (216 slices) from centers 2 and 3—12 patients (53 slices) from center 2 and 18 patients (163 slices) from center 3—was held out for final evaluation only.

Blinding procedures

During generation of the reference scar masks via semi-automated FWHM, the expert reader had no access to any output of the proposed automated pipeline, including model-predicted myocardial contours, scar probability maps, or final automated segmentations, and was further blinded to clinical outcomes. The expert-correction task on the center-1 development subset was performed on a non-overlapping cohort and at an earlier time point, eliminating any potential anchoring effect on the external reference standard.

Enhancement segment annotation

For training the enhancement-probability predictor, the LV myocardium in each LGE slice of the fine-tuning set was divided into 72 chord segments of 5º each. A CMR expert (P.Q.) classified each segment as enhanced or non-enhanced based on the presence of scar tissue, and Gaussian smoothing was applied to these binary labels to obtain soft targets for training.

Pipeline for automated LGE scar quantification

The proposed automated LGE analysis framework is visualized in Figure 2, which includes two main stages as described below.

Stage 1: semi-supervised myocardial segmentation

The myocardial segmentation stage aims to obtain accurate LV epicardial and endocardial contours on LGE images, which serve as the foundation for subsequent scar quantification. Given the scarcity of annotated LGE data, we employ a cross-modal transfer learning strategy combined with expert refinement as illustrated in Figure 2.

Figure 2 Overview of the proposed LGE analysis framework, organized into a development phase (top) and an inference phase (bottom). Development phase. (i) Pseudo masks generation: an nnU-Net trained on cine CMR produces pseudo myocardial masks. (ii) DAFNet: cine and LGE images are separately encoded, aligned via a spatial transformer network, and decoded by a shared segmentor, with pseudo cine labels and adversarial training bridging the cine-LGE domain gap. (iii) Expert review & fine-tuning: an expert corrects an initial subset of LGE masks, which are then used to supervise-fine-tune the segmentor in a single one-off refinement step. (iv) Predictor pre-training: a ResNet34 is trained on 72 chord segments to output Gaussian-smoothed soft labels P (_s1, … , _s72) of per-segment enhancement probability. Inference phase (fully automated). Step 1: DAFNet segments the LGE and the enhancement predictor outputs per-segment probabilities on myocardium. Step 2: segments are split into remote-myocardium (τ_low =0.15) and scar-region candidates (top 5%, τ_high =0.75). Step 3: seed values v_max, v_min are computed as mean intensities within R_seed and R_remote. Step 4: intensities are min–max normalized and an FWHM threshold θ=1/2mean (I_norm [R_seed)] yields the final scar mask and burden. CMR, cardiac magnetic resonance; DAFNet, dual attention-guided fuzzy network; FWHM, full-width at half-maximum; LGE, late gadolinium enhancement; nnU-Net, no-new-net.

An nnU-Net was first trained on the publicly available ACDC dataset to establish a baseline cine segmentation model. Based on the trained nnU-Net, we generated myocardial segmentation masks for in-house cine CMR. We then employed DAFNet to transfer anatomical knowledge from the cine domain to the LGE domain through disentangled representation learning. Specifically, DAFNet decomposes each input image into a modality-invariant anatomical factor and a modality-specific style factor. A shared segmentor operates on the anatomical factors from both modalities, enabling segmentation supervision from labeled cine images to propagate to unlabeled LGE images. A spatial transformer network aligns anatomical factors across modalities, while adversarial training on both segmentation masks and reconstructed images enforces modality-independent anatomical representations. This design allows the model to segment LGE images by leveraging anatomical knowledge learned from cine, without requiring any LGE ground-truth annotations.

To further improve LGE segmentation accuracy, initial myocardial segmentation predictions were reviewed and corrected by a CMR expert (D.A.). All corrected samples were used for supervised fine-tuning of the segmentation model, with no internal training/validation split. This approach balances the efficiency of automated cross-modal adaptation with the precision required for clinical application.

Stage 2: automated scar quantification

Building upon the LV myocardium and blood pool masks obtained in stage 1, we developed a fully automated scar quantification framework based on the FWHM method (11). The key innovation lies in replacing manual reference region selection with an automated, probability-driven approach.

To automate reference region selection, we trained a segment-wise enhancement probability predictor using a ResNet34 backbone (18) with Gaussian-smoothed soft labels derived from expert annotations. The predictor outputs a 72-dimensional probability vector P=[P (s₁), P (s₂), … , P (s₇₂)], where each element represents the likelihood of enhancement within one of 72 uniformly distributed chord segments spanning the LV myocardium.

Based on the predicted probabilities, we identify the remote myocardium region R_remote as segments with minimal enhancement likelihood:

Rremote={si|P(si))<τlow}[1]

where the low probability threshold, τ_low was set to 0.15. These segments are intersected with the myocardial mask from Stage 1, and the mean pixel intensity within this region is computed as v_min. Conversely, segments with P (s_i) > τ_high (τ_high =0.75) are identified as scar candidates; within these regions, the top 5% of pixels by intensity forming a spatially connected component are designated as the scar seed. To establish the upper intensity reference v_max, we compute the mean intensity within a circular region of 10-pixel radius centered at the LV blood pool centroid, thereby avoiding low-signal structures such as papillary muscles. The values of τ_low, τ_high, and the seed-top fraction were selected on the Center-1 development cohort by a coarse grid search maximizing scar-burden ICC against the FWHM reference, and a sensitivity analysis (Table S3) confirmed that results are stable across a wide range of these values.

The LGE image then undergoes linear intensity normalization:

Inorm(x,y)=I(x,y)−vminvminmax[2]

Scar segmentation follows the FWHM criterion, with the threshold θ defined as half of the mean normalized intensity within the scar seed region, denoted as R_seed:

θ=12⋅mean[Inorm(Rseed)][3]

Pixels within the myocardial mask exceeding θ are classified as scar tissue, while those below θ are classified as normal myocardium. For comparison, a supervised 2D U-Net (19) was additionally trained directly on LGE images using the same expert-corrected fine-tuning set, patient-level splits, and pre-processing pipeline, ensuring a fair head-to-head comparison.

Experiment setup and implementation details

All experiments were implemented using PyTorch 1.7 on NVIDIA A100 graphic processing units (GPUs). For the proposed pipeline, nnU-Net pretraining on ACDC followed the default configuration around 400 epochs. DAFNet was trained with a batch size of 8 and an initial learning rate of 1×10⁻⁴ using the Adam optimizer with 800 epochs. Supervised fine-tuning employed a reduced learning rate of 1×10⁻⁵ to prevent catastrophic forgetting and trained about 100 epochs. The ResNet34-based enhancement probability predictor was trained for 600 epochs with a learning rate of 1×10⁻³ and a cosine annealing schedule.

For the supervised U-Net baseline, we adopted a standard four-stage encoder-decoder (channels 32–512) with instance normalization, optimized to jointly segment LV blood pool, myocardium, and scar via a combined Dice and cross-entropy loss. Training used the Adam optimizer with an initial learning rate of 1×10⁻⁴, cosine annealing, batch size 16, and 400 epochs, with standard data augmentations. Input pre-processing and evaluation protocols were identical to those of the proposed pipeline.

Statistical analysis

Continuous variables are presented as mean ± standard deviation. Segmentation performance was evaluated on a per-slice basis using the Dice coefficient, intersection over union, sensitivity, positive predictive value, 95th-percentile Hausdorff distance, and average surface distance. Patient-level agreement between automated and semi-automatic FWHM measurements of scar burden and mean transmurality was assessed using Pearson correlation, intraclass correlation coefficients (ICC[2,1]; two-way random-effects model, absolute agreement, single measure), and Bland–Altman analysis. Pearson correlation tests were two-sided, with P<0.05 considered statistically significant. Analyses were performed in Python 3.10.10 using SciPy 1.11.4.

Results

Zero-shot LGE segmentation

To evaluate the inherent cross-modal generalization capability of DAFNet, we conducted a zero-shot segmentation experiment. The model was trained exclusively with the segmentation masks of cine images in the public dataset, and no LGE segmentation supervision was involved during the model training. The trained DAFNet model was directly applied to the in-house cine and LGE images. As shown in Table 1, due to the domain shift between the public and private datasets, the pretrained DAFNet model achieved modest segmentation performance for both in-house cine (Dice score of 0.82±0.10, mean of myocardium and blood pool) and LGE (Dice of 0.68, mean of myocardium, blood pool and scar) images. This performance gap highlights the need of further fine-tuning.

Performance improvement via fine-tuning

To bridge the domain gap and enhance segmentation accuracy on LGE images, we further finetune the DAFNet using paired cine and LGE data with expert annotations from center 1. Evaluation on the independent test set from the other 2 centers demonstrated substantial improvements across both overlap- and boundary-based metrics following fine-tuning. As detailed in Table 1, on LGE images the mean Dice score increased from 0.68 significantly to 0.77 (an absolute gain of approximately 9 percentage points) and the mean intersection over union (IoU) increased from 0.58 to 0.68, with concurrent improvements in sensitivity (0.62 → 0.72) and positive predictive value (PPV) (0.64 → 0.80). Similar improvements were observed for cine images, with mean Dice increasing from 0.82±0.10 to 0.89±0.02.

Crucially, the boundary-based metrics, which more directly reflect contour quality relevant to downstream wall-thickness and transmurality calculations, showed even larger relative gains than Dice. For the LV myocardium on LGE, 95th percentile Hausdorff distance (HD95) decreased from 4.92±2.78 to 3.43±1.38 mm (a 30.3% reduction) and average surface distance (ASD) decreased from 1.88±0.64 to 1.29±0.40 mm (a 31.4% reduction); the standard deviation of both distance metrics also dropped by roughly half, indicating not only a tighter mean contour but also markedly fewer slice-level outliers. Improvements were even more pronounced for the LV blood pool (HD95: 7.04 → 3.63 mm, −48.4%; ASD: 2.62 → 1.44 mm, −45.0%), which is particularly important because the blood-pool intensity serves as the upper-intensity reference (v_max) in the FWHM step. These results indicate that incorporating expert-annotated data effectively guides DAFNet to learn domain-invariant anatomical features, thereby achieving high-precision segmentation for both cine and LGE modalities on unseen test data. In addition to the quantitative results in Table 1, visual inspection indicates that fine-tuning improves both the delineation of the LV myocardium and the stability of scar localization on LGE as illustrated in Figure 3.

Figure 3 Comparison of myocardial and scar segmentation results across four representative AMI cases (A-D). The rows from top to bottom display: (I) original LGE images; (II) GT established via expert manual correction of predicted segmentations; (III) results from the zero-shot (Pre-FT) pipeline; and (IV) results from the fine-tuned (Post-FT) pipeline. The myocardium is visualized in blue, and scar tissue is highlighted in orange. Quantitative metrics, including the Dice score (relative to GT) and SR, are provided in the bottom-right corner of each panel. AMI, acute myocardial infarction; FT, fine-tuning; GT, ground truth; LGE, late gadolinium enhancement; SR, scar ratio.

Compared with the supervised U-Net baseline trained on the same expert-corrected LGE data (section Experiment setup and implementation details), the proposed fine-tuned pipeline achieved higher performance across all three structures (Table 1). On the three-region mean, it outperformed the U-Net baseline (Dice 0.77 vs. 0.72; HD95 4.50 vs. 6.56 mm; ASD 1.55 vs. 2.40 mm), with the largest gain observed for scar—the most clinically relevant and challenging class—where Dice improved from 0.47 to 0.57 and HD95 decreased from 11.20 to 6.55 mm. The U-Net baseline also exceeded the zero-shot DAFNet on scar Dice (0.47 vs. 0.43), indicating that direct supervision on scarce LGE labels is beneficial but insufficient to match cross-modal transfer combined with expert-corrected fine-tuning.

Scar quantification performance

To assess the clinical utility of the proposed framework, we evaluated the accuracy of automated myocardial scar quantification. Based on the segmentation masks generated by the fine-tuned DAFNet, we applied the automated FWHM algorithm to segment scar tissue. The scar burden, defined as the percentage of scar volume relative to the total myocardial volume, was calculated for each subject in the external test set of center 2 and center 3.

To establish a reference standard, the CMR radiologist (D.A.) performed semi-automated FWHM quantification using ITK-SNAP software (20). This manual procedure involved: (I) defining endocardial and epicardial contours, (II) placing an ROI in the blood pool, (III) selecting scar seeds within hyper-enhanced regions, and (IV) defining the remote myocardium ROI. Scar segmentation was then performed based on the FWHM threshold derived from these manual inputs.

We compared the scar burden derived from our automated pipeline (using both zero-shot and fine-tuned DAFNet masks) against the reference standard. The quantitative agreement between our automated pipeline and the manual reference is illustrated in Figure 4.

Figure 4 Agreement analysis of automated scar quantification against the semi-automated reference standard. (A) Zero-shot model results; (B) fine-tuned model results. Top: scatter plots with Pearson correlation (r), and ICC. Solid red line: regression; dashed black line: identity. Bottom: Bland-Altman plots showing bias (red) and 95% limits of agreement (blue). Fine-tuned model shows reduced bias and narrower limits. FWHM, full-width at half-maximum; ICC, intraclass correlation coefficient; LoA, limits of agreement; SD, standard deviation.

First, regarding scar burden quantification (Figure 4), the domain adaptation strategy yielded a marked improvement in accuracy. The zero-shot baseline exhibited a moderate correlation (r=0.83, P<0.001) but showed a systematic negative bias of −3.73%, indicating a tendency to underestimate scar volume in unseen LGE data. In contrast, the fine-tuned DAFNet achieved high concordance with the radiologist’s measurements, improving the Pearson correlation to 0.93 (P<0.001) and reducing the bias to −3.08% with narrower limits of agreement.

Furthermore, we evaluated the mean transmurality (Figure 5), a clinical index highly sensitive to the precise delineation of myocardial boundaries. The zero-shot model struggled with this task, resulting in a low ICC of 0.66 and a substantial underestimation bias of −11.65%. However, incorporating fine-tuning significantly mitigated this domain gap. The fine-tuned model improved the ICC to 0.835 and the Pearson correlation to 0.88, demonstrating that the proposed framework can reliably replicate expert-level assessment for complex clinical indices. Bland-Altman analysis further confirmed the good agreement between our fully automated method (fine-tuned) and the manual reference, suggesting that on the external test set our approach can replicate expert-level quantification without manual intervention.

Figure 5 Agreement analysis of automated transmurality assessment against the semi-automated reference standard. (A) Zero-shot model results; (B) fine-tuned model results. Top: scatter plots with Pearson correlation (r), and ICC. Solid red line: regression; dashed black line: identity. Bottom: Bland-Altman plots showing bias (red) and 95% limits of agreement (blue). Fine-tuned model shows reduced bias and narrower limits. FWHM, full-width at half-maximum; ICC, intraclass correlation coefficient; LoA, limits of agreement; SD, standard deviation.

To verify that these results do not depend on the specific choice of probability thresholds, we performed a 36-point grid sensitivity analysis (τ_low ∈ {0.10, 0.15, 0.20}, τ_high ∈ {0.70, 0.75, 0.80, 0.85}, seed-top ∈ {3%, 5%, 7%}) on the center-1 development cohort. Across the entire grid, ICC remained ≥0.77, Pearson r≥0.91, and the absolute bias varied by less than 0.4 percentage points, with the default operating point lying on a broad plateau of high agreement (Table S3).

Runtime and hands-on time analysis

We compared operator hands-on time and computational runtime on 10 randomly selected studies from the external test cohort (centers 2/3; 5–8 short-axis LGE slices per study, 70 slices in total). The semi-automated FWHM workflow required 80.89±19.22 s of expert hands-on time per case (range, 58.81–108.16 s), primarily for manual remote-reference ROI placement and contour adjustment. In contrast, our automated pipeline required no user interaction. Inference was performed on a workstation equipped with an NVIDIA A100 GPU, with slices processed sequentially (batch size =1); one-time model loading (~1.7 s) was excluded from per-case timing. The mean GPU processing time was 1.93±0.37 s per study (0.27±0.03s per slice), of which cross-modal cine-LGE segmentation accounted for approximately 94% of compute time and scar quantification for the remainder. Moreover, because all pipeline components are deterministic at inference, repeated runs on identical inputs produced bit-identical outputs (test-retest ICC =1.000 for both scar burden and mean transmurality), in contrast to published inter-observer ICCs of 0.74–0.92 for semi-automated FWHM quantification (13,21).

Discussion

Key findings

This study presented a fully automated framework for myocardial scar quantification from LGE CMR, targeting two major barriers to clinical translation: (I) the scarcity of pixel-level annotated LGE datasets for robust LV myocardium segmentation, and (II) operator-dependent variability in conventional threshold-based scar quantification.

Our key findings are twofold. First, cross-modal transfer from cine to LGE can provide a strong initialization for LGE myocardium segmentation: the zero-shot evaluation showed that a DAFNet model trained only with cine annotations achieved an acceptable Dice score on LGE images, suggesting that modality-invariant anatomical representations can be learned. Importantly, adding limited expert correction and fine-tuning substantially improved boundary accuracy toward clinical-grade performance, indicating that high-quality LGE segmentation can be achieved without exhaustively labeling large LGE cohorts.

Second, the proposed segment-based probability predictor enables automated and reproducible reference region selection for threshold-based scar quantification. By estimating enhancement probability over evenly distributed myocardial segments, the method automatically identifies remote myocardium and scar seed regions, thereby reducing the variability introduced by manual ROI placement (12,13). The high concordance between the automated pipeline and semi-automated quantification supports the feasibility of combining deep learning-based localization with clinically interpretable thresholding (11).

Strengths and limitations

A major strength of this work is its practical balance between annotation efficiency and clinical accuracy. By leveraging cross-modal transfer from annotation-rich cine CMR to the LGE domain and then refining the model with a limited amount of expert-corrected LGE data, the framework reduces the need for exhaustive pixel-level labeling while still improving boundary precision to a level more suitable for downstream pathology quantification. Another strength is that the issues of inter-observer variability is explicitly addressed at two critical steps of the workflow: LV myocardium segmentation and the initialization/reference selection required by threshold-based scar quantification. By automating these components, the pipeline directly targets two dominant sources of inter-operator variation: contouring inconsistency and subjective ROI placement, while preserving interpretability using clinically familiar FWHM-based thresholding rather than relying solely on opaque end-to-end scar segmentation. Published inter-observer reproducibility for FWHM-based scar quantification has reported ICC values typically in the range of 0.74–0.92 for ischemic scar. Our automated pipeline achieved an ICC of 0.840 [95% confidence interval (CI): 0.705–0.917] for scar burden and 0.835 (95% CI: 0.657–0.935) for mean transmurality versus the semi-automated reference. These figures fall within the reported human inter-observer reproducibility range, supporting the claim that the automated pipeline can substitute for the most variability-prone manual steps without exceeding the typical reader-to-reader disagreement. In addition, the modular design allows future upgrades of individual components (e.g., replacing the segmentation backbone or the reference-selection predictor) without rebuilding the entire system, which supports long-term adaptability as imaging protocols and algorithms evolve. Beyond accuracy and reproducibility, the pipeline also offers a substantial workflow advantage: on 10 external-test studies, it reduced per-case operator time from 80.89±19.22 s of expert hands-on interaction (semi-automated FWHM) to fully automated GPU inference of 1.93±0.37 s per study on an NVIDIA A100 (an approximately 42-fold reduction), eliminating manual remote-reference ROI placement while retaining the clinically familiar FWHM-style scar definition. This supports feasibility for high-throughput or batch-processing scenarios.

Several limitations should also be acknowledged. First, the current validation was performed in AMI, and the generalizability of the proposed pipeline to other etiologies such as non-ischemic cardiomyopathy remains to be established. Second, the segment-based probability predictor assumes that the myocardium can be evenly partitioned into 72 uniform sectors; this design choice may be less optimal in the presence of severe ventricular remodeling or highly irregular anatomy, potentially affecting the stability of automated reference selection. Third, although FWHM is widely used and comparatively reproducible, it can underestimate scar burden in regions with MVO, where signal intensity may be reduced rather than enhanced (22). Fourth, the reference masks were produced by a single experienced reader, so our agreement metrics reflect concordance with one semi-automated reference rather than paired inter-reader variability; a prospective two-reader study (with and without pipeline assistance) is planned to directly quantify the reduction in inter-observer variability. Finally, while the study is multicenter, broader evaluation across a wider range of scanner vendors, acquisition protocols, and contrast settings would further strengthen evidence for robustness under real-world heterogeneity.

Comparison with similar research

Prior work on cross-modality cardiac magnetic resonance (MR) segmentation has largely focused on unsupervised domain adaptation via image-level translation [e.g., generative adversarial network (GAN)-based methods (23)] or feature-level alignment (7,14). Image translation approaches may introduce geometric distortions or hallucinated structures, which can be problematic for clinical measurement (24). Feature-level approaches such as DAFNet leverage disentangled representations to separate anatomy from modality and to align anatomical factors across domains. Our results are consistent with this line of research in that cross-modal disentanglement enables meaningful zero-shot transfer to LGE. However, we further show that zero-shot performance alone may be insufficient for clinical-grade boundary accuracy for LGE segmentation, and that incorporating limited expert correction can provide a pragmatic path to improve segmentation without requiring exhaustive LGE labeling. This conclusion is reinforced by the supervised U-Net baseline: despite being trained on the identical expert-corrected LGE data with identical patient-level splits, the U-Net underperformed the proposed pipeline on the three-region mean (Dice 0.72 vs. 0.77) and, most markedly, on scar segmentation (Dice 0.47 vs. 0.57; HD95 11.20 vs. 6.55 mm), supporting the choice of cross-modal transfer over direct supervised learning on scarce LGE annotations.

For scar quantification, conventional practice often relies on semi-automated thresholding methods such as n-SD and FWHM, both of which require manual selection of remote myocardium and/or scar seed ROIs and thus suffer from inter-operator variability. End-to-end learning-based scar segmentation (25-27) has been explored as an alternative, but these methods can be data-hungry and may generalize poorly across centers, and they often provide less direct interpretability than threshold-based criteria. In contrast, our approach aligns with a hybrid strategy: it uses a learned model to standardize the ROI/reference selection step while retaining FWHM for the final delineation. Compared with fully manual or semi-automated workflows, this reduces operator dependence; compared with fully end-to-end scar segmentation networks, it maintains interpretability and leverages clinically established quantification logic.

Explanations of findings

The observed benefit of cross-modal transfer likely arises because LV anatomy (shape and spatial organization) is largely shared across cine and LGE despite substantial appearance differences. Disentangled representation learning facilitates the extraction of modality-invariant anatomical factors, enabling meaningful initialization segmentation of LGE even without LGE labels. Nevertheless, LGE-specific artifacts, low myocardium-blood contrast, and imperfect alignment between modalities can degrade boundary precision which is reflected in the relatively high HD95 (≈4.9 mm for myocardium and ≈7.0 mm for blood pool) and the wide slice-level variance of the zero-shot baseline. This provides a plausible explanation for why expert-corrected fine-tuning yields substantial gains: a small amount of targeted supervision can correct systematic failure modes that unsupervised alignment cannot resolve. Importantly, these boundary-quality gains translate directly into the downstream clinical task: the 30 % reduction in myocardial HD95/ASD and the increase in PPV from 0.711 to 0.824 (i.e., fewer false-positive contour expansions that would otherwise inflate scar volume) coincide with, and provide a mechanistic explanation for, the observed improvement in transmurality ICC from 0.657 (95 % CI: 0.424–0.814) to 0.835 (95 % CI: 0.657–0.935) and in scar-burden ICC from 0.695 to 0.840 on the same images. This concordance reinforces that fine-tuning produces clinically meaningful—rather than merely numerically larger—segmentation improvements.

For scar quantification, intensity heterogeneity in LGE makes absolute thresholds unreliable and increases dependence on where reference ROIs are placed. By turning reference selection into a segment-wise probability estimation problem, the predictor provides a spatially regularized decision mechanism that is less sensitive to local noise and partial-volume effects. In particular, Gaussian-smoothed soft labels encourage learning of gradual enhancement transitions rather than hard class boundaries, which may improve robustness in ambiguous regions and lead to more stable selection of remote myocardium and scar seed segments. The resulting agreement with semi-automated analysis suggests that standardizing the ROI step can meaningfully reduce variability while preserving the clinical rationale of FWHM.

Implications and actions needed

Clinically, a fully automated and reproducible pipeline may facilitate large-scale retrospective analyses and multicenter studies where consistent scar quantification across time points is required. Reducing operator-dependent bias is particularly important for longitudinal monitoring and treatment-response evaluation (4,12,13). The modular design also supports incremental upgrades (e.g., replacing the segmentation backbone or updating the probability predictor) without redesigning the entire workflow.

Several actions are needed before broader deployment. First, the method should be validated on broader disease spectra (including non-ischemic cardiomyopathies) and on more diverse acquisition protocols and vendors. Second, exploring adaptive myocardial sectorization or anatomy-aware partition strategies may improve robustness in remodeled ventricles. Third, incorporating three-dimensional volumetric analysis may better capture transmural extent and spatial heterogeneity of scarring. Finally, uncertainty quantification could provide case-level confidence estimates to support appropriate human oversight, especially in the presence of artifacts or MVO.

Conclusions

We have presented a fully automated framework for myocardial scar quantification from LGE images that addresses two critical challenges: annotation scarcity and operator-dependent variability. By leveraging cross-modal semi-supervised learning, our approach achieves accurate left ventricle myocardial segmentation with minimal annotation burden. The proposed segment-based probability predictor enables automated reference region selection, removing the manual contouring and ROI placement steps that drive observer variability in conventional FWHM workflows. Validation on an external multi-center AMI cohort demonstrated good agreement with expert semi-automated measurements (ICC =0.840 for scar burden, 0.835 for mean transmurality), while reducing operator hands-on time from 80.9±19.2 s to zero per case and yielding perfect test-retest reproducibility (ICC =1.000). These results suggest that the proposed pipeline offers a practical solution for standardizing LGE analysis in both clinical practice and large-scale research studies.

Acknowledgments

The authors thank the Advanced MR Imaging Research Lab (AIRLab) and HPC platform of ShanghaiTech University for their support.

Footnote

Data Sharing Statement: Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0022/dss

Peer Review File: Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0022/prf

Funding: This work was supported in part by the National Natural Science Foundation of China under grant 82572351, in part by the Explorer Program of the Science and Technology Commission of Shanghai Municipality under grant 23TS1400300, and in part by Shanghai QiYuan Innovation Foundation.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://rx.amegroups.com/article/view/10.21037/rx-26-0022/coif). H.Q. serves as an unpaid editorial board member of Radiology-X from February 2026 to January 2028. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Ren Ji Hospital, Shanghai Jiao Tong University School of Medicine (No. [2018]093), and informed consent was waived due to the retrospective nature of this study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Virani SS, Alonso A, Aparicio HJ, et al. Heart Disease and Stroke Statistics-2021 Update: A Report From the American Heart Association. Circulation 2021;143:e254-743. [Crossref] [PubMed]
2. Kim HW, Farzaneh-Far A, Kim RJ. Cardiovascular magnetic resonance in patients with myocardial infarction: current and emerging applications. J Am Coll Cardiol 2009;55:1-16. [Crossref] [PubMed]
3. Wu KC, Zerhouni EA, Judd RM, et al. Prognostic significance of microvascular obstruction by magnetic resonance imaging in patients with acute myocardial infarction. Circulation 1998;97:765-72. [Crossref] [PubMed]
4. Kramer CM, Barkhausen J, Bucciarelli-Ducci C, et al. Standardized cardiovascular magnetic resonance imaging (CMR) protocols: 2020 update. J Cardiovasc Magn Reson 2020;22:17. [Crossref] [PubMed]
5. Bernard O, Lalande A, Zotti C, et al. Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved?. IEEE Trans Med Imaging 2018;37:2514-25. [Crossref] [PubMed]
6. Dou Q, Ouyang C, Chen C, et al. PnP-AdaNet: plug-and-play adversarial domain adaptation network with a benchmark at cross-modality cardiac segmentation. arXiv 2018.
7. Chartsias A, Papanastasiou G, Wang C, et al. Disentangle, Align and Fuse for Multimodal and Semi-Supervised Image Segmentation. IEEE Trans Med Imaging 2021;40:781-92. [Crossref] [PubMed]
8. Wang H, Huang H, Wu J, et al. Semi-supervised segmentation of cardiac chambers from LGE-CMR using feature consistency awareness. BMC Cardiovasc Disord 2024;24:571. [Crossref] [PubMed]
9. Yuan Y, Wang X, Yang X, et al. Semi-supervised class imbalanced deep learning for cardiac MRI segmentation. MICCAI 2023;459:
10. Kellman P, Arai AE. Cardiac imaging techniques for physicians: late enhancement. J Magn Reson Imaging 2012;36:529-42. [Crossref] [PubMed]
11. Amado LC, Gerber BL, Gupta SN, et al. Accurate and objective infarct sizing by contrast-enhanced magnetic resonance imaging in a canine myocardial infarction model. J Am Coll Cardiol 2004;44:2383-9. [Crossref] [PubMed]
12. Heiberg E, Engblom H, Carlsson M, et al. Infarct quantification with cardiovascular magnetic resonance using "standard deviation from remote" is unreliable: validation in multi-centre multi-vendor data. J Cardiovasc Magn Reson 2022;24:53. [Crossref] [PubMed]
13. Flett AS, Hasleton J, Cook C, et al. Evaluation of techniques for the quantification of myocardial scar of differing etiology using cardiac magnetic resonance. JACC Cardiovasc Imaging 2011;4:150-6. [Crossref] [PubMed]
14. Chartsias A, Joyce T, Papanastasiou G, et al. Disentangled representation learning in cardiac image analysis. Med Image Anal 2019;58:101535. [Crossref] [PubMed]
15. Liu X, Sanchez P, Thermos S, et al. Learning disentangled representations in the imaging domain. Med Image Anal 2022;80:102516. [Crossref] [PubMed]
16. Qi H, Qian P, Tang L, et al. Predicting Late Gadolinium Enhancement of Acute Myocardial Infarction in Contrast-Free Cardiac Cine MRI Using Deep Generative Learning. Circ Cardiovasc Imaging 2024;17:e016786. [Crossref] [PubMed]
17. Isensee F, Petersen J, Klein A, et al. nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation. arXiv 2018.
18. He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition. arXiv 2015.
19. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv 2015.
20. Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 2006;31:1116-28. [Crossref] [PubMed]
21. Gavrysh J, Reisdorf P, Hadler T, et al. Reproducibility of late gadolinium enhancement quantification techniques in ischemic and non-ischemic heart diseases (ReLate study). Front Cardiovasc Med 2025;12:1621292. [Crossref] [PubMed]
22. Bulluck H, Rosmini S, Abdel-Gadir A, et al. Impact of microvascular obstruction on semiautomated techniques for quantifying acute and chronic myocardial infarction by cardiovascular magnetic resonance. Open Heart 2016;3:e000535. [Crossref] [PubMed]
23. Zhu JY, Park T, Isola P, et al. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. arXiv 2020.
24. Cohen JP, Luck M, Honari S. Distribution matching losses can hallucinate features in medical image translation. arXiv 2018.
25. Zabihollahy F, White JA, Ukwatta E. Convolutional neural network-based approach for segmentation of left ventricle myocardial scar from 3D late gadolinium enhancement MR images. Med Phys 2019;46:1740-51. [Crossref] [PubMed]
26. Fahmy AS, Neisius U, Chan RH, et al. Three-dimensional Deep Convolutional Neural Networks for Automated Myocardial Scar Quantification in Hypertrophic Cardiomyopathy: A Multicenter Multivendor Study. Radiology 2020;294:52-60. [Crossref] [PubMed]
27. Tavakoli N, Rahsepar AA, Benefield BC, et al. ScarNet: a novel foundation model for automated myocardial scar quantification from late gadolinium-enhancement images. J Cardiovasc Magn Reson 2025;27:101945. [Crossref] [PubMed]