MFDL Best Papers

MFDL Best Papers – First Half of 2026

Han, T. H., Kim, Y. W., Lee, H. J., Kim, J. S., Lee, S. G., Yang, D. H., Oh, H. M., Kim, D., Shin, S. Y., Song, S., Lee, J. S.* "Synthesis of coronary 4D CT Image by denoising diffusion probabilistic model", Computer Methods and Programs in Biomedicine (2026)

This study proposes a diffusion-based framework for time-variant hemodynamic analysis in coronary artery disease (CAD). Conventional CT-derived fractional flow reserve (CT-FFR) relies on single-phase CT and does not account for coronary motion, whereas full 10-phase 4D CT entails a cumulative radiation dose of approximately 150 mSv. Fluid-structure interaction (FSI) analysis also requires vessel mechanical properties that cannot be measured in individual patients. The framework synthesizes 4D CT images from the end-diastolic and end-systolic phases alone, using a denoising diffusion probabilistic model (DDPM) and a deformation module that estimates vessel deformation from the images. Pressure drop, FFR, and wall shear stress are then computed across the cardiac cycle by computational fluid dynamics (CFD) with quasi-steady FSI. The synthesized images achieved a peak signal-to-noise ratio of 32.01, a structural similarity index measure of 0.937, and a mean Dice coefficient of 0.973 after segmentation. The 4D CT-FFR showed a diagnostic accuracy of 90.5% compared with invasive FFR, and the two-phase protocol reduced radiation dose by 80% relative to full 10-phase acquisition.

MFDL Best Papers – Second Half of 2025

Kim, S. Y., Kwon, S. W., Bashir, M. N., Lee, J. S.* "Predicting dendrite growth in lithium metal batteries through iterative neural networks and voltage embedding", npj Computational Materials (2025)

This study addresses the computational challenges of predicting lithium dendrite growth during repeated charge–discharge cycles by developing a deep learning framework that integrates a one-dimensional convolutional neural network (1D CNN), iterative training, and physics-based voltage embedding (VE). Conventional molecular dynamics (MD) simulations provide atomistic insights into dendrite formation and interfacial reactions but require substantial computational resources, while existing machine learning approaches primarily focus on macroscopic battery properties or short-term predictions. In contrast, the proposed framework enables long-term forecasting of atomic positions, charge distributions, voltage propagation, and dendritic morphology across multiple cycles. The model achieves a mean atomic position error of 1.53%, maintains high structural accuracy (Pearson correlation > 0.96), and accurately reproduces dendrite morphologies under different electrolyte conditions (Dice coefficients of 0.92–0.94). Furthermore, voltage embedding reduces the voltage prediction error from 8.68% to 4.80%, while computation time decreases from 18 hours to 25 minutes compared with conventional MD simulations. These results demonstrate a substantial advancement in efficient atomistic forecasting, offering a practical approach for investigating battery degradation mechanisms and accelerating electrolyte screening and next-generation battery development.

MFDL Best Papers – First Half of 2025

Kwon, S. W., Moon, J. K., Song, S. C., Cha, J. Y., Kim, Y. W., Choi, Y. J., Lee, J. S.*. "Time-series X-ray image prediction of dental skeleton treatment progress via neural networks", Computers in Biology and Medicine (2025).

This study addresses the challenges in accurately predicting skeletal changes during orthodontic treatment in growing patients by utilizing advanced generative neural networks, specifically denoising diffusion probabilistic models (DDPM), latent diffusion models (LDM), and ControlNet. Traditional approaches, such as support vector regression and multilayer perceptrons, rely heavily on multiple sequential radiographs, increasing radiation exposure and landmark identification errors without providing visually interpretable predictions. In contrast, the diffusion-based methods explored in this research demonstrate superior predictive accuracy and clinical practicality with minimal input images, notably achieving comparable performance with a single-input LDM and ControlNet. These methods significantly reduce clinical requirements, patient visits, and radiation exposure, offering personalized, visually interpretable orthodontic treatment plans based on patient-specific attributes, thereby marking a substantial advancement in orthodontic predictive modeling.