AI Revolutionizes Cardiac and Blood Flow MRI Imaging

Abstract

Dynamic MRI reconstruction, one of inverse problems, has seen a surge by the use of deep learning techniques. Especially, the practical difficulty of obtaining ground truth data has led to the emergence of unsupervised learning approaches. A recent promising method among them is implicit neural representation (INR), which defines the data as a continuous function that maps coordinate values to the corresponding signal values. This allows for filling in missing information only with incomplete measurements and solving the inverse problem effectively. Nevertheless, previous works incorporating this method have faced drawbacks such as long optimization time and the need for extensive hyperparameter tuning. To address these issues, we propose Dynamic-Aware INR (DA-INR), an INR-based model for dynamic MRI reconstruction that captures the spatial and temporal continuity of dynamic MRI data in the image domain and explicitly incorporates the temporal redundancy of the data into the model structure. As a result, DA-INR outperforms other models in reconstruction quality even at extreme undersampling ratios while significantly reducing optimization time and requiring minimal hyperparameter tuning.

A groundbreaking medical artificial intelligence (AI) technique now enables the precise and high-quality reconstruction of MRI images even from incomplete scan data. This innovative approach not only shortens reconstruction times compared to existing methods but also offers greater ease of use for medical professionals, promising to improve diagnostic accuracy in clinical settings.

Developed by Professor Jaejun Yoo and his research team from the UNIST Artificial Intelligence Graduate School (AIGS), the new AI model-Dynamic-Aware Implicit Neural Representation (DA-INR)-advances the field of dynamic MRI reconstruction.

Dynamic MRI captures fast-changing physiological signals such as heartbeat and blood flow, making it invaluable for diagnosing various conditions. However, traditional imaging approaches face limitations, including lengthy scan durations and the challenge of acquiring full datasets within a single session. Consequently, reconstructing high-resolution images from partial, incomplete data has become a critical area of research.

DA-INR addresses these challenges by overcoming the complex parameter tuning and lengthy optimization times common in previous models. Inspired by the fact that core tissue structures-such as atria and ventricles of the heart or blood vessels in the liver-remain consistent over time, the team conceptualized these as a unified "canonical space." Instead of reconstructing each frame independently, DA-INR models the static tissue structures within this canonical space and reflects temporal changes relative to it. This approach significantly reduces unnecessary computations and minimizes noise and distortions that typically occur during frame-by-frame reconstruction.

The results are impressive: DA-INR outperforms existing leading models in both image quality and efficiency. The Peak Signal-to-Noise Ratio (PSNR), an indicator of image sharpness, improved by up to 1.5 dB, while Structural Similarity Index (SSIM), reflecting structural fidelity, increased by over 0.01. Moreover, reconstruction time was reduced by more than sevenfold, and memory usage was cut by more than 50%.

Beyond sharper images, DA-INR accurately captures physiological motion, such as the rapid contraction and slower relaxation of the heart-overcoming the common 'Over-Smoothing' problem that hampers traditional AI models. This advancement has been demonstrated in dynamic contrast-enhanced (DCE) liver scans, where the model effectively differentiates between healthy tissue and lesions, such as liver tumors, by capturing their characteristic rapid contrast changes.

Professor Yoo commented, "This technology allows for high-resolution spatiotemporal MRI reconstruction using only limited scan sequences, without the need for additional data. Its simplicity means that medical practitioners can readily adopt it without extensive technical adjustments."

This research was led by Dayoung Baik as the first author and will be presented at the upcoming MICCAI 2025, one of the most prominent international conferences in medical AI. The study was supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP), the National Research Foundation of Korea (NRF), and UNIST.

Journal Reference

Dayoung Baik and Jaejun Yoo, "Dynamic-Aware Spatio-temporal Representation Learning for Dynamic MRI Reconstruction," '25 MICCAI, (2025).