Work

Manifold learning and inverse processes for multimodal superresolution in T1 and T2 weighted magnetic resonance imaging

Year: 2026
Type: article
Abstract: We propose a novel multimodal superresolution algorithm based on the analysis of graph Laplacian representations of image data. The algorithm is inspired by a model for heterogeneous data fusion. We demonstrate its application to the problem of magnetic resonance imaging, where superresolution can be used to decrease acquisition times or to improve signal-to-noise ratio. At the core of the proposed algorithm is a joint modality feature matching, performed in the Laplacian eigenmaps embedding space. A sparse L1-regularized preimage reconstruction is then required, in order to produce a high resolution image from its graphical features. To optimize the algorithm's performance, we integrate the spatial information of the image data to improve local correlations. To illustrate the performance of this new method we apply it to the problem of reconstructing high-resolution T1-weighted magnetic resonance images from a combination of high resolution T2-weighted and low resolution T1-weighted images. Our results indicate that the improvement, as compared to conventional single image interpolation schemes, is proportionally greater for low-resolution images. Our results support the use of the manifold learning embedding approach to multimodal superresolution, by demonstrating superior numerical performance over state-of-the-art image superresolution algorithms.
Source: PLoS ONE
Language: English
Open Access status: gold
Estimated APC paid: $2,382
OpenAlex ID: W7236123034
FWCI:
Cites: 41
Cited by:
Related to: 9
Parts: 12