Accurate and early stroke lesion segmentation is necessary for better results for patients and successful treatment. A novel deep learning model using a multi-modality architecture that combines spatial, channel, and temporal information from perfusion MRI and CT images is proposed in this work for the deep segmentation of stroke lesions. The model addresses cross-domain generalization by integrating dynamic perfusion information from the ISLES 2022 dataset, such as Diffusion Weighted Imaging (DWI) and CT Perfusion (CTP) images alongside high-resolution anatomical data obtained from the ATLAS v2.0 dataset. We enhance the proposed approach by adding a hybrid spatial channel temporal attention transformer block to a UNet encoder-decoder architecture. Multi-scale features that are patch-embedded and enhanced with positional encoding are extracted by the encoder. To increase sensitivity to anatomical features and corresponding pathological changes, the attention module integrates temporal encoding of perfusion parameters like Cerebral Blood Flow (CBF), Cerebral Blood Volume (CBV), Mean Transit Time (MTT), and Time-To-Maximum (Tmax) with joint modeling of spatial and channel dependencies. The experimental results contribute to reliable clinical application by exhibiting increased accuracy and robustness over traditional UNet and unimodal transformer models.