AI Reads Heart Motion From CT Scans to Detect Hidden Heart Attacks

Heart attacks leave a fingerprint that no static scan can fully capture: the injured heart muscle moves differently. A team of Japanese researchers has now built an artificial intelligence system that reads that fingerprint directly from four-dimensional cardiac CT scans, tracking the deformation of the beating myocardium frame by frame and using the resulting motion fields to flag myocardial infarction. The framework, described in the International Journal of Computer Assisted Radiology and Surgery, reached an accuracy of 91.92 percent and an area under the ROC curve of 0.9335 on an in-house clinical dataset, outperforming conventional feature-based machine learning baselines while revealing a subtler and arguably more important result: adding motion information did not dramatically improve overall discrimination, but it consistently lifted the model’s sensitivity to true infarction cases.

The system, named MambaMorph–Motion Infarction Detector, or MMMID, was developed by Ryosuke Sasaki, Yusuke Kurose and colleagues working with Tatsuya Harada at RIKEN and collaborating institutions, together with a clinical team from a partner hospital. Their motivation is grounded in a hard clinical reality. Cardiovascular diseases remain the leading cause of mortality worldwide, and myocardial infarction is among the most critical conditions to detect because it is characterized by abnormal myocardial motion. Diagnosing it properly requires detailed motion analysis, yet assessing sequential motion images is time-consuming and imposes a considerable burden on physicians. Most previous automated approaches have relied on local motion descriptors or limited shape features, such as regional wall shortening, which the authors argue does not fully represent the complex spatiotemporal nature of myocardial movement throughout the cardiac cycle.

MMMID’s pipeline is built from three interconnected deep learning modules. The first performs semantic segmentation, extracting the left ventricle, right ventricle and left ventricular myocardium from CT volumes. The researchers used a 3D U-Net with an encoder–decoder structure and skip connections, initialized with ModelsGenesis pretrained weights, a transfer learning strategy designed to make the segmentation robust to images acquired from different devices. The second module estimates motion: a deformable registration network computes dense spatial correspondences between cardiac phases, producing deformation fields that describe how every point of the myocardium moves from end-diastole, when the heart is most relaxed, toward end-systole, when it is most contracted. The third module is a 3D DenseNet-121 classifier, implemented in MONAI, that takes both the raw CT volumes and the estimated motion fields and predicts whether a patient has myocardial infarction.

The registration engine is where the framework departs from earlier work. Traditional deformable registration optimizes an energy function to align a moving image with a fixed image, while learning-based methods such as VoxelMorph and TransMorph use neural networks to estimate deformation quickly. MMMID instead employs MambaMorph, which replaces the Swin Transformer block in TransMorph with a Mamba block built on a state-space model. In a Mamba block, the input is first projected through MLP and 1D convolution layers, and the computations then pass to the SSM layer, which performs sequence-to-sequence mapping through a hidden state updated by transition matrices that are discretized via zero-order hold. The crucial property is computational: state-space models handle long sequential data with linear complexity, making them faster than Transformer-based alternatives while still capturing long-range dependencies. For volumetric cardiac motion estimation, where both spatial context and temporal consistency matter, that combination is advantageous.

Training the registration module required no explicit ground-truth motion annotations. The model was optimized with a weak-supervised Dice loss, which rewards anatomical alignment between the deformed moving volume and the fixed volume, combined with a smoothness regularization on the deformation field weighted at 0.1 across all experiments. In practice, MambaMorph takes a concatenated pair of target and fixed volumes, flows the input through two branches, one processing the full volume and a UNet-like branch splitting it into tokens for the Mamba block, and outputs the deformation field. When the researchers evaluated the resulting motion fields by deforming end-diastolic segmentation masks toward end-systole and comparing them with ground-truth labels, the mean Dice similarity coefficients were high, and expert physicians confirmed the tracking accuracy was sufficient. In a direct comparison under the same Dice-based protocol, MambaMorph generally yielded higher Dice scores than VoxelMorph across cardiac phases and anatomical regions.

The classification stage is where anatomy and motion are fused. The researchers defined SeqMotion as a temporal stack of three-channel displacement fields, covering the x, y and z components, estimated at 10 percent intervals of the cardiac cycle from end-diastole at 0 percent to end-systole at 40 percent, with each field masked to the left ventricular myocardium using the segmentation output. The final classifier input concatenates the paired end-diastolic and end-systolic CT volumes with this SeqMotion stack along the channel dimension, producing a multi-channel 3D tensor that encodes both static anatomy and dynamic contraction patterns. The classification dataset contained 266 normal cases and 81 myocardial infarction cases collected between January 4 and December 27, 2022, reflecting the natural clinical distribution rather than an artificially balanced sample. All scans came from a 256-slice Philips scanner with retrospective ECG-gated reconstruction achieving a temporal resolution of approximately 27 milliseconds per sector.

Evaluated with stratified fivefold cross-validation at the patient level, with a fixed operating point set by the argmax of the two output logits rather than any test-set threshold optimization, MMMID achieved 91.92 percent accuracy and an AUC of 0.9335. It surpassed an adapted LVSnake-style baseline, which divided the left ventricular myocardium into 27 regions and fed pooled displacement statistics into a support vector machine, and it beat an 88-feature radiomic approach, which extracted shape and texture descriptors from the left ventricular cavity and surrounding myocardium at both key phases, in accuracy and F1-score. A DeLong test found no statistically significant difference in AUC between MMMID and the radiomics-based method, with a p-value of 0.77, so the headline numbers should be read as competitive rather than decisive.

The ablation studies tell the more nuanced story. A model using only raw CT volumes achieved an AUC of 0.9340, essentially identical to the 0.9335 achieved when sequential motion fields were added, and the paired DeLong test showed no statistically significant difference between the CT-plus-SeqMotion and CT-only configurations, with p equal to 0.41. But recall, the proportion of infarction cases correctly detected, rose from 71.47 percent with CT-only input to 77.65 percent with CT-plus-SeqMotion input. Sequential motion fields also improved classification accuracy by 2.9 percent compared with using a single aggregated motion field between end-diastole and end-systole, and the CT-plus-SeqMotion configuration achieved a significantly higher AUC than SeqMotion alone, with p below 0.001, while the difference against CT-plus-Motion was not significant at p equal to 0.346. The authors interpret this as evidence that CT appearance already provides strong discriminative information, and that estimated deformation fields alone are insufficient for robust detection, because infarction is reflected not only in motion but also in structural changes such as myocardial thinning visible on CT.

Qualitative case analysis reinforced the complementary role of motion. In successful detections, the model captured reduced myocardial motion in the central region of the left ventricle and along the endocardial border during mid-systole, in patients whose expert clinical assessment showed decreased ejection fraction. In failed cases, relatively active central ventricular motion during mid-systole may have confounded the estimation, and in one notable failure the physician’s report localized the infarction to the inferior wall, a region inherently difficult to assess with the current CT slice-based input configuration. The researchers caution that the estimated motion fields provide useful qualitative information but should not be regarded as validated infarct localization, since the framework performs case-level classification only.

The limitations are stated plainly. The dataset came from a single institution, external validation was not performed, and generalizability to different scanners, acquisition protocols and patient populations remains uncertain. The classification dataset was imbalanced, with fewer infarction cases than normal cases, so performance metrics should be interpreted with that imbalance in mind. The authors also note that further systematic comparisons with other recent registration architectures such as TransMorph, deformation plausibility analyses including smoothness and folding evaluation, and attention to error propagation from segmentation and registration into the final classification all remain necessary. Future work will focus on multicenter external validation, improved fusion between CT volumes and motion fields, regional infarct localization when detailed annotations become available, and prospective clinical evaluation to determine whether the framework can genuinely reduce physicians’ diagnostic burden in cardiac CT analysis.

Subject of Research: Deep learning detection of myocardial infarction from cardiac CT using deformable registration-based myocardial motion estimation

Article Title: Myocardial infarction detection framework using cardiac motion estimation with deformable registration field

Article References: Sasaki, R., Kurose, Y., Iho, J., Tokunaga, Y., Horie, M., Nishizawa, K., Hayashi, Y., Koyama, Y., & Harada, T. (2026). Myocardial infarction detection framework using cardiac motion estimation with deformable registration field. International Journal of Computer Assisted Radiology and Surgery. https://doi.org/10.1007/s11548-026-03797-8

Image Credits: AI Generated

DOI: 10.1007/s11548-026-03797-8

Keywords: myocardial infarction, cardiac CT, deformable image registration, MambaMorph, state-space model, deep learning, medical imaging, myocardial motion, DenseNet, heart disease detection, 4D imaging, computer-aided diagnosis