Automatic Diagnosis and Subtyping of Ischemic Stroke Based on a Multidimensional Deep Learning System

The classification and subtyping of ischemic stroke play a significant role in therapeutic decision-making, which leads to higher requirements on the infarct lesion location. This study established a novel deep learning system (DLS) integrating four different models, blood supply, border zone, brain structure atlas, and cortex, to provide a multidimensional delineation of ischemic stroke lesions. Based on the atlas segmentation logic of each model, the output channels of the convolutional neural networks were rearranged, and the image preprocessing parameters were optimized to achieve faster convergence speed. The proposed multidimensional DLS is applied to the magnetic resonance diffusion weighted imaging (MRI-DWI) image examinations to accurately locate the infarction lesions in four specific types of territories. Two experienced neuroradiologists evaluate the performance of the system. About 90% of the results are acceptable in clinical evaluation tests, proving the established system’s feasibility in comprehensive ischemic stroke classification and subtyping and its further potential in diagnosing other brain disorders.