arXiv subjects

Elmira Mirzabeigi

Publications and source records attributed to Elmira Mirzabeigi.

4 recordsLinked to original sources

arXiv2026

A$^3$PreCore: Robust Rigid Registration for Native FLAIR-Space Brain-Mask Transfer

Reliable brain-mask transfer into native fluid-attenuated inversion recovery (FLAIR) space can fail when T1-to-FLAIR rigid registration is unstable. We developed A$^3$PreCore, a six-degree-of-freedom rigid-registration strategy that aligns the full bias-corrected T1-weighted image to native FLAIR using normalized mutual information while preserving the established T1-space anatomical mask. We also introduced extended quantitative quality control (E-QC), which characterizes mask extent, field-of-view-relative position, FLAIR intensity support, boundary support, inferior morphology, and pairwise geometric disagreement. In 78 participants with paired T1-weighted and FLAIR imaging, A$^3$PreCore was compared with fsl_anat and the Brain Extraction Tool (BET). Gross rigid-body misregistration occurred in 21/78 fsl_anat outputs (26.9%) and 0/78 A$^3$PreCore outputs (exact two-sided McNemar p=9.54 x 10^-7). Among 57 participants without gross fsl_anat misregistration, A$^3$PreCore preserved close agreement (median Dice=0.993; median paired Delta NMI=-0.00007, p=0.429). A$^3$PreCore remained broadly concordant with BET (median Dice=0.955), with less inferior and extracranial extension in representative comparisons. Method-blinded neuroscientists ranked A$^3$PreCore first in 35/40 participants (87.5%). E-QC additionally exposed severe orientation-related disagreement not evident from global descriptors or centroid displacement alone. These findings support A$^3$PreCore as a robust native-FLAIR brain-mask transfer strategy and E-QC as a complementary framework for multidimensional mask characterization and participant-specific geometric quality control.

cs.CV↗

arXiv2025

BridgeNet: A Hybrid, Physics-Informed Machine Learning Framework for Solving High-Dimensional Fokker-Planck Equations

BridgeNet is a novel hybrid framework that integrates convolutional neural networks with physics-informed neural networks to efficiently solve non-linear, high-dimensional Fokker-Planck equations (FPEs). Traditional PINNs, which typically rely on fully connected architectures, often struggle to capture complex spatial hierarchies and enforce intricate boundary conditions. In contrast, BridgeNet leverages adaptive CNN layers for effective local feature extraction and incorporates a dynamically weighted loss function that rigorously enforces physical constraints. Extensive numerical experiments across various test cases demonstrate that BridgeNet not only achieves significantly lower error metrics and faster convergence compared to conventional PINN approaches but also maintains robust stability in high-dimensional settings. This work represents a substantial advancement in computational physics, offering a scalable and accurate solution methodology with promising applications in fields ranging from financial mathematics to complex system dynamics.

physics.comp-ph↗

arXiv2025

LyAm: Robust Non-Convex Optimization for Stable Learning in Noisy Environments

Training deep neural networks, particularly in computer vision tasks, often suffers from noisy gradients and unstable convergence, which hinder performance and generalization. In this paper, we propose LyAm, a novel optimizer that integrates Adam's adaptive moment estimation with Lyapunov-based stability mechanisms. LyAm dynamically adjusts the learning rate using Lyapunov stability theory to enhance convergence robustness and mitigate training noise. We provide a rigorous theoretical framework proving the convergence guarantees of LyAm in complex, non-convex settings. Extensive experiments on like as CIFAR-10 and CIFAR-100 show that LyAm consistently outperforms state-of-the-art optimizers in terms of accuracy, convergence speed, and stability, establishing it as a strong candidate for robust deep learning optimization.

cs.LG↗

arXiv2023

Designing of knowledge-based potentials via B-spline basis functions for native proteins detection

Knowledge-based potentials were developed to investigate the differentiation of native structures from their decoy sets. This work presents the construction of two different distance-dependent potential energy functions based on two fundamental assumptions using mathematical modeling. Here, a model was developed using basic mathematical methods, and the carbon-alpha form is the simplest form of protein representation. We aimed to reduce computational volume and distinguish the native structure from the decoy structures. For this purpose, according to Anfinsens dogma, we assumed that the energy of each model structure should be more favorable than the corresponding native type. In the second one, we thought that the energy difference between the native and decoy structures changes linearly with the root-mean-square deviation of structures. These knowledge-based potentials are expressed by the B-spline basis functions of the pairwise distances between Ca-Ca of inter-residues. The potential function parameters in the above two approaches were optimized using the linear programming algorithm on a massive collection of Titan-HRD and tested on the remainder. We found that the potential functions produced by Anfinsens dogma detect native structures more accurately than those developed by the root-mean-square deviation. Both linear programming knowledge-based potentials (LPKP) successfully detect the native structures from an ensemble of decoys. However, the LPKP of the first approach can correctly identify 130 native structures out of 150 tested cases with an average rank of 1.67. While the second approach LPKP detects 124 native structures from their decoys. We concluded that linear programming is promising for generating knowledge-based potential functions. All the training and testing sets used for this work are available online and downloaded from http://titan.princeton.edu/HRDecoys.

math.OC↗