Work

Improving the accuracy of perfusion map generation using dynamic susceptibility contrast magnetic resonance imaging data based on recurrent neural networks

Year: 2026
Type: article
Abstract: This study explores the process of generating perfusion parameter maps from time series of DSC-MRI images of the brain. The task addressed relates to the inaccuracy and inefficiency of perfusion map generation due to insufficient analysis of temporal features, dependence on the arterial inflow function, as well as excessive computational complexity of the models. This paper proposes an approach to generating perfusion parameter maps using neural networks that directly process time series of perfusion images. Three deep learning architectures have been designed and experimentally investigated, differing in the way in which temporal information is taken into account and the presence of recurrent layers. The evaluation was performed on an open medical data set with patient-level sample distribution to prevent information leakage between the training and test parts. A controlled comparison of models with and without recurrent layers was performed to determine the impact of explicitly taking into account temporal dependences on the accuracy of map generation. The normalized root mean square error decreased from 0.027 to 0.016 and 0.015. The structural similarity index increased from 0.826 to 0.957 and 0.973. The peak signal-to-noise ratio increased from 31.493 to 36.095 and 36.412. Additional comparison to approaches reported in other studies showed that the proposed architecture with recurrent layers demonstrates competitive or higher values ​​of image quality metrics. The results confirm the feasibility of using neural networks with recurrent layers for more accurate generation of perfusion parameter maps. The practical significance is the possibility of integrating the approach into automated perfusion analysis systems and clinical decision support in the diagnosis of stroke and brain tumors
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $242
OpenAlex ID: W7166802406
FWCI: 4.856
Cites: 29
Cited by: 2
Related to: 9