A novel MRI image-analyzed stroke prediction framework using multiscale residual long short-term memory with spatial attention network

In developed countries; stroke is reported as the second leading reason of mortality. The quick clinical intervention and examination have an important impact on the patient’s life quality and also prevent the growth of infarct. By performing mechanical or pharmaceutical interventions, the clinical interventions concentrate on restoring the perfusion deficits. The clinicians are required to analyze the merits and risks based on multimodal neuroimaging examinations including their own clinical experience and Magnetic Resonance Imaging (MRI) scans, regardless of which reperfusion are utilized. This decision-making operation could improve from an automatic final infarct prediction that would offer a tissue evaluation. Hence, in this work, a unique stroke prediction framework is designed by considering deep learning approaches. At first, an essential MRI images are gathered from publicly available datasets. Hence, the gathered images are fed to the segmentation phase by developed Adaptive Multi-path Refinement Networks (AMRN). Moreover, the segmentation accuracy of the model is strengthened by tuning network parameters using Statistical American Zebra Optimization (SAZO). Later, the segmented images are given into the stage of feature extraction to extract the deep features using Graph Convolutional Neural Network (GCNN), texture feature, and morphological feature from the segmented images. Further, the extracted features are given into the designed Multiscale Residual Long Short-Term Memory with Spatial Attention Network (MResLSTM-SAN) for predicting the stroke. Finally, the outcome of the recommended technique is validated with a set of classical models.