Deep Learning-Based Multi-Feature Auxiliary Diagnosis Method for Early Detection of Ischemic Stroke

Abstract

Stroke, an acute cerebrovascular disease, has become the second leading cause of death worldwide after coronary heart disease, characterized by high incidence, disability, and mortality rates, with an increasingly younger affected population. Clinically, stroke types are primarily divided into ischemic and hemorrhagic strokes, with ischemic stroke being the most common. Presently, early identification methods rely heavily on physicians' experience, leading to misdiagnosis, missed diagnosis, diagnostic delay, and other issues, potentially resulting in worsened conditions or severe complications. Although artificial intelligence-based stroke auxiliary diagnosis systems have been employed in recent years to reduce missed diagnoses and enhance work efficiency, their impact on improving diagnostic accuracy has been limited. The main reason for this limitation is the selection of relatively singular or atypical feature types in neural networks, which affects diagnostic accuracy. To address this issue, this study leverages the rapid and sensitive response of electroencephalogram (EEG) data to cerebral ischemia and combines it with clinical indicators to propose a comprehensive "clinical indicators + quantitative electroencephalogram" multi-feature pattern recognition method. Initially, 23 key features for neural network training are selected. Subsequently, an ischemic stroke diagnosis model combining LSTM attention and multi-feature is constructed. In an experiment involving 500 ischemic stroke patients, the diagnostic model demonstrates an accuracy of 0.81, a sensitivity of 0.82, and an F1-score of 0.81. Moreover, to accurately locate the lesion area, the three-dimensional features of MRI images are used. A cascaded 3D deep residual network stroke precise segmentation method is constructed by incorporating residual units and cascade concepts into the 3DCNN network. The evaluation indicators of this segmentation algorithm on the training set are: DICE coefficient 0.91, precision 0.94, and sensitivity 0.89. Experimental results indicate that the proposed method outperforms existing clinical diagnosis schemes and CNN segmentation models in terms of diagnostic performance. The implementation of rapid and accurate diagnosis during early stages of stroke onset is crucial for improving ischemic prognosis, minimizing brain damage, and reducing mortality and disability rates.