Abstract
Stroke, the second leading cause of death globally, necessitates prompt diagnosis for effective prognosis. CT imaging has limitations, especially in identifying acute lesions. This work introduces a novel deep repre sentation that uses multimodal inputs from CT studies and perfusion parametric maps, to retrieve stroke lesions. The architecture follows an autoencoder representation that forces attention on the geometry of stroke through additive cross-attention modules. Besides, a cascade train is herein proposed to generate synthetic perfusion maps that complement multimodal inputs, refining stroke lesion segmentation at each stage of processing and supporting the observational expert analysis. The proposed approach was validated on the ISLES 2018 dataset with 92 studies; the method outperforms classical techniques with a Dice score of .66 and a precision of .67.