Spatiotemporal deep learning for scar screening in cardiovascular magnetic resonance: Toward selective use of gadolinium Authors: Amine Amyar, Shiro Nakamori, Jiwon Kim, Martin S Maron, Ethan J Rowin, Karishma Pareek, Alexander Schulz, Robert M Judd, Warren J Manning, Raymond Y Kwong, Frederick L Ruberg, Jonathan W Weinsaft, Reza Nezafat Published: 2026 ABSTRACT AbstractBackgroundLate gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) is the gold standard for assessing myocardial scar. However, a substantial proportion of patients referred for CMR scar assessment are ultimately found to have no evidence of myocardial scarring.PurposeTo develop and evaluate a deep learning (DL) model capable of identifying patients without myocardial scar using cine imaging alone, thereby obviating the need for contrast administration and LGE imaging in these individuals.Materials and MethodsWe developed a novel spatiotemporal DL architecture to identify patients unlikely to have myocardial scar using contrast-free cine images. The model was trained on short-axis cine images from a consecutive cohort of 3000 patients (1753 males; mean age 54 ± 18 years) undergoing CMR for evaluation of known or suspected cardiovascular disease, using 1.5 and 3T scanners from Siemens and GE. External validation was performed in an independent multicenter cohort of 1792 patients where images were acquired on 1.5T and 3T scanners from Siemens and Philips. The architecture utilizes factorized convolutions to extract spatial and temporal features and incorporates residual attention mechanisms to emphasize features most predictive of scar presence on LGE imaging. To evaluate the incremental value of incorporating temporal information, a second model was developed that excluded the temporal kernel from the DL architecture. Both models were trained and optimized using the same training dataset and were evaluated based on similar internal and external testing cohorts. Model performance in identifying patients without myocardial scar was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.ResultsThe spatiotemporal model had a higher AUC compared to the spatial model in the internal (0.79 ± 0.02 vs. 0.70 ± 0.05, p<0.05, n = 500) and external cohort (0.78 vs. 0.64, p<0.001, n = 1792). The spatiotemporal model correctly identified 64% (176/274) and 52% (410/789) of patients without scar in the internal and external test sets, while maintaining a high sensitivity (86% and 82%).ConclusionsIncorporating temporal information from cine images using an end-to-end spatiotemporal DL architecture enables non-contrast screening for myocardial scar.Keywords: Contrast-fee CMR, Myocardial scar screening, Gadolinium, Spatiotemporal neural network Introduction Cardiovascular diseases are the leading cause of death globally [1]. Cardiovascular magnetic resonance (CMR) is widely regarded as the non-invasive imaging gold standard for evaluating cardiac structure and function [2], [3]. Late gadolinium enhancement (LGE) on CMR is the gold standard for myocardial scar assessment and requires the injection of a gadolinium-based contrast agent (GBCA) [4], [5]. While gadolinium is generally considered safe, concerns persist regarding tissue deposition [6], [7], [8], allergic and other adverse reactions [9], [10], periodic global shortages [11], prolonged scan times, the increased healthcare costs [12], and environmental contamination [13], [14]. Furthermore, prior studies have reported that approximately 50% of clinically referred patients show no LGE on CMR [15], [16]. Therefore, there is an unmet need to reduce or eliminate the use of GBCA in clinical practice. In recent years, deep learning (DL)-based techniques have emerged as promising tools to reduce the use of GBCA by generating synthetic LGE-like images from non-enhanced short-axis balanced steady-state-free-precession cine images [17], [18], [19], [20] or in combination with native T1 mapping [19], [21]. For instance, Zhang et al. developed a conditional generative adversarial network to create a synthetic LGE from cine images and native T1 maps for myocardial infarction assessment [21]. The model used two separate encoder-decoder branches, one for each input modality, which were then fused through a third encoder-decoder to generate the synthetic LGE. This architecture was also trained on hypertrophic cardiomyopathy patients with a third branch to include T1-weighted images [22]. More recently, another study investigated the use of cine images alone as input to a generative model that extracts spatial and temporal features to detect scar regions, leveraging style transfer learning to synthesize LGE-like images in patients with acute myocardial infarction [20]. However, despite the potential of synthetic LGE, these approaches still require substantial development on large datasets to cover the wide spectrum of cardiovascular diseases, as well as thorough validation, before they can be adopted in clinical practice. Rather than entirely replacing LGE with non-contrast methods, an alternative approach is to minimize GBCA use to selectively reserve gadolinium administration for patients with a high likelihood of myocardial scar when referred for scar evaluation. Recent studies have explored this approach by combining non-contrast imaging with radiomics to detect the presence of myocardial fibrosis in hypertrophic cardiomyopathy (HCM) patients. For instance, Mancio et al. developed a machine learning model using radiomic features extracted from cine images to predict the presence of myocardial fibrosis [23]. Similarly, Fahmy et al. proposed an approach that combines radiomics with a convolutional neural network to detect myocardial scarring from cine images [15]. However, these methods do not account for the effects of scar on myocardial motion dynamics throughout the cardiac cycle, which often accompany the presence of scar. In this study, we developed an end-to-end spatiotemporal neural network to extract both textural (spatial) and temporal features from cine images for myocardial scar screening. We hypothesize that incorporating temporal features of wall motion would improve scar detection compared with texture-based features alone. Materials and methods Study population and CMR image acquisition All data were collected retrospectively from six different medical centers (Fig. 1). The institutional review boards in the six centers approved the use of CMR data for research with a consent waiver. Patient information was handled in compliance with the Health Insurance Portability and Accountability Act. CMR data for model development and internal testing were acquired at Weill Cornell Medical Center (Cornell) between 2016 and 2020 using 1.5T and 3T scanners from GE Healthcare (Chicago, Illinois) and Siemens Healthineers (Erlangen, Germany). The dataset initially consisted of 3236 consecutive patients referred for evaluation of known or suspected cardiovascular disease. The study protocol included balanced steady-state free-precession cine imaging acquired before gadolinium administration and LGE. Images were reviewed by an experienced reader (S.N., >10 years of experience). A total of 236 patients were excluded due to non-diagnostic LGE (severe artifacts) or severe motion artifacts on cine images, resulting in a final cohort of 3000 patients (1753 males; mean age 54 ± 18 years). LV function measurements and the presence of scar were obtained from the corresponding clinical reports as interpreted by clinical readers as part of routine practice. Training and testing strategy. The spatiotemporal deep learning model was developed using data from a single center, collected with Siemens (1.5 and 3T) and GE scanners (1.5T and 3T) (A). The data were split into training (n = 2000), validation (n = 500), and internal testing (n = 500) in five-fold cross-validation setting. The model with the best AUC curve from the internal testing was then externally evaluated on data from five different centers (n = 1792). The external testing data were collected using Siemens (3T) and Philips (1.5T) (B) to evaluate model generalizability. LGE late gadolinium enhancement, AUC area under the receiver operating characteristic curve To assess the generalizability of the model, we performed external validation using data collected from five additional medical centers (Table S2): Beth Israel Deaconess Medical Center (BIDMC, n = 296, 145 LGE+), Tufts Medical Center (Tufts, n = 628, 362 LGE+), Boston Medical Center (BMC, n = 388, 212 LGE+), Brigham and Women’s Hospital (BWH, n = 214, 73 LGE+), and Mie University Hospital (Mie, n = 134, 79 LGE+). Patients were referred for CMR for evaluation of known or suspected cardiovascular disease. Images were acquired between 2004 and 2021 on 1.5T and 3T scanners from Siemens and Philips (Best, the Netherlands). All anonymized DICOM images were transferred to BIDMC. Pre-contrast ECG-gated breath-hold cine images in short-axis view were collected with the following sequence parameters: generalized autocalibrating partially parallel acquisition rate of 2, TR = 2.8–3.7 ms, TE = 1.1–1.5 ms, flip angle = 41°–60°, pixel resolution = 1.1–1.7 mm, slice thickness = 6.2–10 mm, and bandwidth = 580–1334 Hz. All images were centrally cropped to 128 × 128 pixels and normalized to an intensity range of 0 to 1. Data about biventricular function and volume were extracted from the clinical report. Short-and long-axis 2D inversion-recovery LGE images were acquired using a breath-hold, segmented inversion-recovery sequence with the following parameters: 6–8 mm slice thickness, 10-mm inter-slice gap, TR, 6.4 ms; TE, 1.2–2.7 ms; FA, 25°; and spatial resolution of 1.7–2.0 mm2. Deep learning architecture We developed a custom convolutional neural network (CNN) architecture to identify myocardial scar from contrast-free cine CMR images (Fig. 2), comprising a spatial-only model and an extended spatiotemporal model. The spatial CNN served as a baseline and was designed to capture spatial patterns within individual cine frames without modeling temporal dependencies. In this model, image frames were organized along the channel dimension and processed using three-dimensional spatial convolutions combined with spatial attention, enabling the network to learn frame-specific anatomical features. Each frame was processed independently, thereby restricting the learned representation to spatial information only. Spatiotemporal deep learning model architecture for scar screening: The neural network architecture comprises a series of factorized convolutions, followed by an attention module, layer normalization, projection, and downsampling, forming a residual attention block (A). The factorized convolution applies a 3D convolution across the spatial dimensions, followed by a 1D convolution across the temporal dimension (B). The final output is the probability of a patient having a myocardial scar. For comparison, a spatial neural network without the temporal component was developed. Each frame was processed independently, without accounting for temporal dependencies across frames. 1D one-dimensional 3D three-dimensional Building upon this baseline, we developed a spatiotemporal CNN that explicitly incorporates temporal dynamics across the cardiac cycle. The model's input is a 4D volume consisting of a stack of 10 short-axis cine images (a 3D volume with 20 frames representing the temporal dimension). In cases where the number of slices or frames was less than the expected input size, bicubic interpolation was used. Directly processing a 4D volume in a neural network is highly challenging due to the lack of support for 4D convolutions in existing DL libraries and the significant computational memory burden when learning directly from such high-dimensional data [24], [25], [26]. To address this, we developed a 4D (3D+time) layer that decomposes the 4D information into spatial (3D) and temporal (1D) components. This was achieved by factorizing the 4D kernel: first, applying a 3D spatial convolution of size 3 × 3 × 3 to extract spatial hierarchical features, followed by a 1D temporal convolution of size 11 × 1 × 1 to capture temporal features. The large temporal kernel is necessary to maintain long-range dependencies across the 20 frames. Given the large volume of extracted features, we apply spatiotemporal attention using a SoftMax activation function to highlight spatial and temporal regions most indicative of scar presence. Regions highly relevant to scar presence receive scores close to 1, while scar-free regions receive scores near 0. This factorization of spatial and temporal kernels reduces the number of model parameters from 46 million to 1.4 million while preserving rich spatiotemporal feature representations. Following the factorized convolution layer, we implement four successive residual attention blocks [27]. Each block consists of a series of attention modules, followed by a factorized convolution, layer normalization to enhance robustness against batch size variations and mitigate the vanishing gradient problem [28], and a projection layer without an activation function to project the final feature maps back to the input dimension for concatenation. This projection was necessary to form a residual connection, which effectively addresses the vanishing gradient issue during backpropagation [29]. Additionally, a downsampling is applied to prevent overfitting and manage the large-scale input data. Applying spatiotemporal attention after factorized convolution enables the extraction of global heart motion while capturing long-range dependencies between frames. Additionally, the residual attention block extracts both textural and temporal features at multiple scales, enabling the detection of regional motion and subtle changes that may indicate the presence of scar. The output of the fourth residual attention block undergoes global average pooling instead of traditional fully connected layers. This enforces the relationship between the learned feature maps and the probability of scar presence while significantly reducing the number of model parameters, thereby lowering the risk of overfitting. The final layer is fully connected with a single neuron and a sigmoid activation function, outputting the probability of the patient having a myocardial scar. Model training Five-fold cross-validation was employed to split the development and internal testing dataset into development (2500 patients) and internal testing (500 patients). Stratified patient-wise random sampling was used to maintain the same number of patients with and without scar in the training and testing subsets. The development subset was further split into training (2000 patients) and validation (500 patients) subsets for determining the optimal model parameters. The testing subset is referred to as the internal testing set because it is sampled from the same cohort used for model development. The model was optimized using mini-batch stochastic gradient descent with a batch size of 64, an initial learning rate of 0.001, and a weight decay of 0.0001, with adjustments made when validation loss plateaued. Training was conducted for up to 500 epochs, with early stopping set at 70 epochs to prevent overfitting. Binary cross-entropy (BCE) was used as the loss function, as the task is a binary classification problem (scar vs. no scar). BCE quantifies the difference between the predicted probability of the positive class and the true label, penalizing predictions that deviate from the ground truth. The loss is averaged across all samples during training. BCE is particularly suitable for probabilistic outputs, as it encourages the model to predict probabilities close to 1 for positive cases and close to 0 for negative cases. Balanced binary accuracy was monitored during training and validation to ensure model stability. All models were implemented in TensorFlow 2.4.1 and trained on an NVIDIA DGX-1 system with eight Tesla V100 GPUs (NVIDIA, Santa Clara, California), each with 32 GB of memory and 5120 cores. All code for training and testing is freely available online (https://github.com/HMS-CardiacMR/NonGad_scar_screening). Statistical analysis Continuous variables are expressed as mean ± standard deviation and categorical variables are reported as counts and percentages. For a fair comparison, both spatial and spatiotemporal models were trained, optimized, and tested on the same sets. For each fold of the five-fold cross-validation, the definition of the optimal cutoff point was determined using the Youden index. A patient was classified as LGE+ if the predicted probability of scar, P(Scar | optimal_cutoff), exceeded the optimal cutoff on the development dataset. The area under the ROC curve (AUC) was calculated for both models to predict the presence of a scar. The DeLong test was used to compare AUCs and subsequently adjusted for multiple comparisons using the false discovery rate procedure according to Benjamini-Hochberg [30]. Sensitivity represented the portion of patients with myocardial scar who were correctly identified by the model. Specificity represented the proportion of patients without myocardial scar who were correctly identified by the model, thereby potentially avoiding unnecessary administration of GBCA. Positive predictive value (PPV), defined as the proportion of true positives among all positive predictions, was calculated to assess the reliability of positive test results, while negative predictive value (NPV), defined as the proportion of true negatives among all negative predictions, was calculated to assess the reliability of negative test results. All data analysis was performed using Python 3.9.6. Statistical analyses were conducted using the Python libraries scikit-learn (version 0.19.1) and scipy-stats (version 1.12.0; Python Software Foundation, 2024). The model with best AUC computed on the internal testing dataset was used as the final model applied to the external testing dataset. Results Baseline characteristics of study patients Clinical and demographic data for the internal and external cohorts are summarized in Table 1, Table 2, and S1. In the development and internal testing cohort, the mean LV end-diastolic volume index, LV end-systolic volume index, LV mass index, and LV ejection fraction were 79 ± 24, 33 ± 20, 56 ± 16, and 66 ± 21, respectively. In the external cohort, corresponding values were 82 ± 39, 45 ± 32, 57 ± 17, and 75 ± 15, respectively. LV ejection fraction and LV mass index showed a good separation between patients with and without LGE on the development and internal testing cohort (AUC of 0.70 ± 0.03 and 0.72 ± 0.02, respectively), while this decreased to an AUC of 0.57 ± 0.01 and 0.65 ± 0.03 in the external dataset (supplementary material). Table 1Patient’s characteristics for the development and internal testing datasetAll patients (n = 3000)Development cohort (n = 2500)Internal testing cohort (n = 500)AllLGE+ (n = 1130)LGE- (n = 1870)AllLGE+ (n = 904)LGE- (n = 1596)AllLGE+ (n = 226)LGE- (n = 274)Age, years56±1861±1652±1856±1862±1652±1857±1760±1554±19Male, %1753 (58%)806 (71%)947 (51%)1434 (57%)645 (71%)789 (49%)319 (64%)161 (71%)158 (58%)BSA1.90±0.261.95±0.261.88±0.261.90±0.261.95±0.261.88±0.261.91±0.261.91±0.251.90±0.27LV EDVI, ml/m279±2485±3177±2078±2383±3077±2082±2693±3577±19LV ESVI, ml/m233±2042±3030±1333±1940±2930±1436±2450±3629±12LV EF, %56±1648±1760±1156±1549±1760±1154±1645±1761±10LV mass index, g/m266±2180±2761±1666±2179±2661±1670±2586±3263±18Data is represented as n (%) or mean ± deviation. BSA body surface area, LV left ventricle, EDVI end diastolic volume index, ESVI end systolic volume index, EF ejection fraction, LGE late gadolinium enhancement Patient’s characteristics for the development and internal testing dataset Data is represented as n (%) or mean ± deviation. BSA body surface area, LV left ventricle, EDVI end diastolic volume index, ESVI end systolic volume index, EF ejection fraction, LGE late gadolinium enhancement Table 2Patient’s characteristics for the external datasetAllLGE+ (n = 1003)LGE- (n = 789)Age, years53±1655±1651±16Male, %1085 (61%)661 (66%)424 (54%)BSA1.95±0.271.94±0.271.95±0.28LV EDVI, ml/m282±3986±4078±37LV ESVI, ml/m245±3248±3542±29LV EF, %57±1755±1860±15LV mass index, g/m275±1581±2968±23Data is represented as n (%) or mean ± standard deviation. BSA body surface area, LV left ventricle, EDVI end diastolic volume index, ESVI end systolic volume index, EF ejection fraction, LGE late gadolinium enhancement Patient’s characteristics for the external dataset Data is represented as n (%) or mean ± standard deviation. BSA body surface area, LV left ventricle, EDVI end diastolic volume index, ESVI end systolic volume index, EF ejection fraction, LGE late gadolinium enhancement Model training Fig. 3 shows the training and validation loss curves for both the spatiotemporal and spatial models. In both cases, the losses decreased progressively and reached stability, with minimal differences observed between training and validation. For the spatiotemporal model, the BCE loss remained below 0.1 for both training and validation, indicating excellent learning performance. For the spatial model, the loss was below 0.2, reflecting very good training performance. Convergence was achieved before the maximum of 500 epochs, stabilizing around 350 epochs for the spatiotemporal model and 400 epochs for the spatial model, with early stopping applied when no further improvement was observed. Training and optimization of each fold required approximately three days on average, while inference was completed in less than 1 s per patient on a single V100 GPU (NVIDIA, Santa Clara, California). Binary cross-entropy loss curves for the spatiotemporal and spatial models during training and validation. All losses decreased steadily across epochs until reaching stability around epoch 350 for the spatiotemporal model and 400 for the spatial model, at which point training was terminated by early stopping after 70 consecutive epochs without improvement. The final loss values for both training and validation remained below 0.1 for spatiotemporal, and 0.2 for the spatial model, indicating excellent model training Myocardial scar screening with spatiotemporal network The spatiotemporal network consistently outperformed the spatial-only network across internal and external testing datasets. In the internal testing dataset, the spatiotemporal model achieved a significantly higher AUC compared to the spatial network (0.79 ± 0.02 vs. 0.70 ± 0.05; p<0.05). It also demonstrated superior sensitivity (0.86 ± 0.06 vs. 0.63 ± 0.06) while maintaining similar specificity (0.64 ± 0.10 vs. 0.69 ± 0.09), indicating better discrimination between patients with and without LGE (Table 3). PPV was also higher with the spatiotemporal network (0.85 ± 0.05) compared to the spatial model (0.64 ± 0.06), while NPV was higher in the spatial network (0.67 ± 0.05 and 0.72 ± 0.02). Table 3Performance of the scar prediction models trained using 5 different splits of the development datasetCross-validation12345Mean±SDSpatialAUC0.700.610.710.730.730.70±0.05Sensitivity0.690.610.540.630.690.63±0.06Specificity0.660.580.830.730.670.69±0.09PPV0.630.550.720.660.630.64±0.06NPV0.720.640.690.710.720.72±0.02SpatiotemporalAUC0.770.770.820.790.800.79±0.02Sensitivity0.800.930.810.830.910.86±0.06Specificity0.660.520.780.670.580.64±0.10PPV0.800.900.830.830.890.85±0.05NPV0.660.620.750.680.640.67±0.05Performance is evaluated using the internal testing dataset. SD standard deviation, PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve Performance of the scar prediction models trained using 5 different splits of the development dataset Performance is evaluated using the internal testing dataset. SD standard deviation, PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve In the external testing dataset, the spatiotemporal model demonstrated superior performance compared to the spatial model (Table 4). The spatiotemporal network achieved an AUC of 0.78 compared to 0.64 for the spatial model (p<0.001), with higher sensitivity (0.82 vs. 0.63) and comparable specificity (0.59 vs. 0.52), and superior predictive values (PPV 0.70 vs. 0.56, NPV 0.68 vs. 0.66). Table 4Performance of the final spatial and spatiotemporal models for detecting scar from cine images in the external cohort (n = 1792)AUCSensitivitySpecificityPPVNPVSpatial0.640.630.590.560.66Spatiotemporal0.780.820.520.700.68PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve Performance of the final spatial and spatiotemporal models for detecting scar from cine images in the external cohort (n = 1792) PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve Fig. 4 illustrates representative examples of model performance across scenarios of success and failure. The model successfully detected myocardial scar from non-enhanced short-axis cine images alone (Fig. 4A), relying on changes in texture and wall motion. It also correctly identified patients without scar (Fig. 4D). Misclassifications were frequently observed in patients with absent or only mild wall motion abnormalities (WMA) despite the presence of scar (Fig. 4B), or in patients without scar but with pronounced WMA (Fig. 4C). Examples from four patients illustrating model performance: (A) correct prediction of scar from non-contrast short-axis cine images, (B) failure to predict scar presence, (C) false-positive prediction of scar, and (D) correct prediction of scar absence. In these cases, scar predictions appeared to depend on the presence or absence of wall motion abnormalities, with accurate prediction of scar absence when such abnormalities were lacking. Myocardial displacement is shown on the cine images using feature tracking from cvi42, visualized by circles and their displacement between end-systole and end-diastole. Yellow arrows highlight regions of wall motion abnormalities, while white arrows indicate the location of scar. LGE late gadolinium enhancement Discussion In this study, we developed and evaluated a spatiotemporal DL model for myocardial scar screening using cine images. Leveraging a large, multicenter, multi-vendor dataset of 4792 patients, the proposed model can identify patients with a low likelihood of myocardial scarring, thereby supporting a triage strategy in which LGE imaging can be selectively reserved for higher-risk patients. Such an approach has the potential to optimize clinical workflow, reduce healthcare costs, and limit unnecessary exposure to gadolinium-based contrast agents, particularly in patients undergoing repeated CMR examinations or those with renal dysfunction. Standard CMR protocols routinely combine cine imaging, for the assessment of cardiac morphology and function, with LGE imaging, the reference standard for myocardial scar detection and tissue characterization. LGE imaging is particularly valuable for quantifying myocardial scar burden and assessing prognosis. Nevertheless, a significant proportion of patients exhibit no detectable fibrosis on LGE imaging, yet continue to undergo repeated gadolinium-based CMR throughout their lifetime [15], [16]. Signal intensity from balanced steady-state-free precession cine images alone is insufficient for reliable scar detection, due to the nonlinear relationship between signal intensity and tissue properties, as well as variations caused by through-plane motion and B1 inhomogeneity [15]. Moreover, distinguishing subtle differences between healthy and scarred myocardium can be challenging in the presence of image noise and artifacts. The presence of a scar can cause WMA that are captured by cine imaging. However, this association is stronger in ischemic cardiomyopathy than in non-ischemic. In a study by Csecs et al. [31], no association was found between LV mechanical deformation parameters and the presence or extent of fibrosis in non-ischemic cardiomyopathy, highlighting the complexity of the relationship between scar burden and LV mechanics. Prior studies have also shown that subtle LV systolic dysfunction precedes regional WMA, reductions in LV ejection fraction, and the development of myocardial scar [32], [33]. To address these challenges, we developed a spatiotemporal DL model capable of learning both textural and temporal features from cine images, capturing multi-scale motion abnormalities and subtle changes in texture that may reflect underlying myocardial scarring. By incorporating residual attention mechanisms, the model can identify both global and regional dysfunction patterns, enhancing its ability to detect scar-related motion abnormalities. Recent years have witnessed growing interest in developing non-contrast imaging techniques for myocardial tissue characterization, with several DL approaches showing promise in detecting fibrosis. However, LGE remains the cornerstone for qualitative myocardial scar and fibrosis assessment. It excels in identifying focal fibrosis and helps predict ischemic damage reversibility [5], distinguishing ischemic from non-ischemic disease [34] and elucidating cardiomyopathy etiology [34], [35]. Fully replacing LGE with non-contrast imaging and DL requires extensive validation across various cardiomyopathies where LGE is indicated, and sensitivity must match that of LGE. Rather than seeking to replace LGE imaging entirely, our model functions as a triage tool, helping to identify patients unlikely to benefit from contrast administration and allowing for a more targeted use of LGE in those with a high likelihood of myocardial scarring. Advances in imaging techniques such as T1 mapping have enabled the quantification of various tissue-specific myocardial properties, broadening the scope of tissue characterization. For instance, areas of scar have a longer T1 than healthy tissues [36], [37]. Nevertheless, T1 mapping is not a substitute for LGE imaging [38], as T1 signal may be influenced by tissue components other than scar, such as fat deposition. Future research should investigate whether integrating T1 mapping with cine imaging can further improve non-contrast scar detection. While in our study, we developed a 4D spatiotemporal convolutional network, transformer-based architectures also offer the potential to capture dynamic information through attention mechanisms across spatial and temporal dimensions. Transformers can, in principle, model long-range dependencies and complex temporal patterns more flexibly than standard convolutional approaches. However, implementing a full 4D transformer for short-axis cine images presents significant challenges, including high computational cost and substantial memory requirements [39]. Further studies are needed to compare our proposed approach head-to-head with a 4D transformer. Despite advances in non-contrast imaging and DL, several challenges remain in implementing a non-contrast screening approach. Previous studies have frequently excluded slices affected by misregistration, inconsistent labeling of ambiguous enhancement, or poor cine image quality. Additionally, one of the largest barriers to implementing a non-contrast approach lies in patient selection [38]. Current studies focus on ischemic patients, where scar presence can be translated into WMA captured by cine imaging. Patients with scar but no WMA, as well as those with non-ischemic cardiomyopathies, are usually excluded. Including these patients is necessary for the integration of the model into clinical practice. In our study, the LV ejection fraction and LV mass index demonstrated good discrimination between patients with and without scar in the development and internal testing datasets; however, this performance decreased significantly in the external testing dataset. The spatial model exhibited a similar trend, whereas the spatiotemporal model maintained high performance across both datasets. This suggests that the spatiotemporal model may be capturing motion information that reflects the presence of scar beyond pixel-wise information. Further studies are warranted to explicitly extract motion features and incorporate them into model training for scar prediction. Finally, previous studies often relied on the interpretation of a single reader to define ground truth for training and validation, which can limit generalizability. In our study, scar presence or absence was determined as part of routine clinical care, with each site’s independent reader reviewing LGE images during clinical interpretation. This approach mirrors real-world workflows and emphasizes the importance of developing a DL model using a large, heterogeneous, multicenter dataset to ensure robust generalizability. Limitations In this study, model development and validation were limited to short-axis cine CMR images. While this view provides comprehensive coverage of LV anatomy and function, the exclusion of long-axis views such as the 2-chamber and 4-chamber planes may have limited the model’s ability to capture additional spatial and functional information. Incorporating multi-planar cine images in future optimization of the model could enhance its performance by offering a more complete representation of myocardial motion and deformation. Furthermore, the inclusion of additional sequences such as T1 mapping may further refine the model’s ability to detect subtle tissue alterations associated with myocardial fibrosis. However, T1 mapping data were not available in the majority of our cases or were limited to 1–3 slices. A multimodal learning approach that integrates both structural and tissue characterization data could improve the model performance. Finally, this was a retrospective study, and future prospective studies are warranted to validate the effectiveness of this screening strategy. Quantitative stratification by scar extent or transmurality was not performed, as measurements of scar burden were not available across the multicenter cohort. Reliable assessment of scar extent would require consistent LGE scar segmentation which was not feasible within the scope of this retrospective study. Conclusion Incorporating temporal information from cine images using an end-to-end spatiotemporal DL architecture enhances non-contrast screening for myocardial scar. Validated in a large multicenter, multi-vendor cohort, this approach may provide a pathway to triage patients for LGE imaging, thereby reducing unnecessary gadolinium administration. Funding Reza Nezafat receives grant funding from the National Institutes of Health (NIH) 1R01HL129185, 1R01HL129157, 1R01HL127015, and 1R01HL154744 (Bethesda, MD, USA). Amine Amyar receives grant funding from the American Heart Association (AHA) 24CDA1257995. Alexander Schulz receives funding from the German Research Foundation (Walter Benjamin Fellowship). Author contributions Amine Amyar: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Shiro Nakamori: Writing – review & editing, Visualization, Methodology, Investigation, Formal analysis, Data curation. Jiwon Kim: Writing – review & editing, Visualization, Formal analysis, Data curation. Martin S. Maron: Writing – review & editing, Visualization, Formal analysis, Data curation. Ethan J. Rowin: Writing – review & editing, Visualization, Formal analysis, Data curation. Karishma Pareek: Writing – review & editing, Visualization, Data curation. Alexander Schulz: Writing – review & editing, Visualization, Formal analysis. Robert M. Judd: Writing – review & editing, Data curation. Warren J. Manning: Writing – review & editing, Supervision, Investigation, Formal analysis, Data curation. Raymond Y. Kwong: Writing – review & editing, Visualization, Formal analysis, Data curation. Frederick L. Ruberg: Writing – review & editing, Visualization, Formal analysis, Data curation. Jonathan W. Weinsaft: Writing – review & editing, Visualization, Formal analysis, Data curation. Reza Nezafat: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Disclosures Drs. Nezafat and Manning have a research agreement with Siemens Healthineers, the manufacturer of MRI system used in a subset of patients for imaging. Declaration of Competing Interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Amine Amyar reports financial support was provided by American Heart Association Inc. Alexander Schulz reports financial support was provided by German Research Foundation. Reza Nezafat reports financial support was provided by National Institutes of Health. The remaining authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Supplementary data associated with this article can be found in the online version at doi:10.1016/j.jocmr.2026.102730. Supplementary material Supplementary material . References Associated Data Supplementary Materials Supplementary material