<p>According to the World Health Organization, the stroke burden is rising in the world. Timely quantification of stroke severity is essential to improve stroke outcomes. Recently, deep learning-based segmentation algorithms have emerged to aid clinicians. However, deep learning techniques require a substantial amount of data for training. Earlier studies encountered challenges with insufficient data for network training. To address this challenge, we developed a model incorporating an intra-domain transfer learning (intra-DTL) framework for multiple modalities and the channel-spatial attention (CSA) module for effective feature extraction. This intra-DTL employs a bottleneck-based multimodal framework that leverages similar data from the same domain. This method integrates common (present in both datasets) and uncommon (not present in both datasets) modalities to enhance feature extraction, enabling effective knowledge transfer. Moreover, the CSA module employs a squeeze-and-excitation module, and region-specific global attention (RSGA) is utilized collectively to enhance the features. Additionally, we incorporated a class-balanced approach during training with the source dataset. Further, we also analyzed the effect of groups in RSGA, different frameworks for effective knowledge transfer, and weight variations analysis. The experiments are carried out on the ISLES 2015 dataset. These experimental results demonstrate that incorporating a bottleneck-based intra-DTL framework, attention, and a class-balanced training approach improved results, achieving a mean Dice score of 0.837. These findings highlight the effectiveness of the intra-DTL framework for multimodal input in transferring knowledge from the source to the target dataset.</p>