Institution profile

Institut PRISME

Academic institutioneurope · fr
Official website
Research library2linked papers
Opportunities0open roles
Selected work

Representative Papers

Robust Sparse Subspace Tracking from Corrupted Data Observations

Sep 20, 2025

This work addresses the problem of robust dynamic subspace estimation and tracking in high-dimensional data streams corrupted by non-Gaussian noise and sparse outliers. To tackle this, we propose a novel online algorithm grounded in α-divergence minimization. By incorporating α-divergence into the subspace optimization objective, the method explicitly enhances robustness against heavy-tailed noise and impulsive data corruptions. Further, it integrates sparse subspace modeling with a low-rank iterative update scheme, ensuring theoretical convergence while substantially reducing computational and memory complexity. Extensive experiments on both synthetic and real-world datasets demonstrate that the proposed approach achieves significantly lower subspace tracking error and higher direction-of-arrival (DOA) estimation accuracy compared to state-of-the-art robust PCA and adaptive subspace methods. These results validate its effectiveness and practicality for dynamic signal processing under challenging non-Gaussian and sparse corruption conditions.

0 citationsRead paper

Feasibility study for reconstruction of knee MRI from one corresponding X-ray via CNN

Mar 16, 2025

This study addresses the high cost and limited accessibility of MRI in early diagnosis of knee osteoarthritis (KOA). We propose a novel end-to-end method for reconstructing 3D MRI volumes from a single X-ray image—without requiring paired X-ray–MRI training data. Our approach innovatively transfers latent-space features learned from X-rays to the MRI synthesis task: a convolutional autoencoder extracts discriminative X-ray representations, and a cross-modal mapping network directly generates anatomically plausible 3D MRI volumes. By eliminating reliance on scarce and costly paired multimodal datasets, our method substantially lowers data acquisition barriers. Evaluated on clinical data, the reconstructed MRIs demonstrate anatomical fidelity and quantitative accuracy, achieving a PSNR of 28.3 dB. This work establishes a feasible, low-cost alternative to conventional MRI and introduces a new paradigm for modality-agnostic medical image synthesis in resource-constrained settings.

0 citationsRead paper
Recent publications

Latest Papers

Robust Sparse Subspace Tracking from Corrupted Data Observations

Sep 20, 2025

This work addresses the problem of robust dynamic subspace estimation and tracking in high-dimensional data streams corrupted by non-Gaussian noise and sparse outliers. To tackle this, we propose a novel online algorithm grounded in α-divergence minimization. By incorporating α-divergence into the subspace optimization objective, the method explicitly enhances robustness against heavy-tailed noise and impulsive data corruptions. Further, it integrates sparse subspace modeling with a low-rank iterative update scheme, ensuring theoretical convergence while substantially reducing computational and memory complexity. Extensive experiments on both synthetic and real-world datasets demonstrate that the proposed approach achieves significantly lower subspace tracking error and higher direction-of-arrival (DOA) estimation accuracy compared to state-of-the-art robust PCA and adaptive subspace methods. These results validate its effectiveness and practicality for dynamic signal processing under challenging non-Gaussian and sparse corruption conditions.

0 citationsRead paper

Feasibility study for reconstruction of knee MRI from one corresponding X-ray via CNN

Mar 16, 2025

This study addresses the high cost and limited accessibility of MRI in early diagnosis of knee osteoarthritis (KOA). We propose a novel end-to-end method for reconstructing 3D MRI volumes from a single X-ray image—without requiring paired X-ray–MRI training data. Our approach innovatively transfers latent-space features learned from X-rays to the MRI synthesis task: a convolutional autoencoder extracts discriminative X-ray representations, and a cross-modal mapping network directly generates anatomically plausible 3D MRI volumes. By eliminating reliance on scarce and costly paired multimodal datasets, our method substantially lowers data acquisition barriers. Evaluated on clinical data, the reconstructed MRIs demonstrate anatomical fidelity and quantitative accuracy, achieving a PSNR of 28.3 dB. This work establishes a feasible, low-cost alternative to conventional MRI and introduces a new paradigm for modality-agnostic medical image synthesis in resource-constrained settings.

0 citationsRead paper