Background: Stroke is a leading cause of morbidity and mortality worldwide. Accurate segmentation of sub-acute and chronic stroke lesions using MRI is crucial for assessing brain damage and developing effective rehabilitation plans. Manual segmentation is labor-intensive and error-prone, necessitating automated approaches. This study aims at improving sub-acute and chronic stroke lesion segmentation using deep learning and multi-modal MRI data. Both models are made available to the research community. Methods: This study developed and evaluated two models for segmenting sub-acute and chronic stroke lesions using MRI: a single-modality model trained on the public ATLAS v2.0 dataset, and a dual-modality model adapted from the single-modality model by integrating T1-w and FLAIR MRI data from an internal dataset. Both models were trained using the nnU-Net framework, employing a preprocessing pipeline to improve the segmentation accuracy. Results: The single-modality model achieved a mean Dice score of 83.0% on the ATLAS v2.0 dataset, and 68.8% on the internal test set. The dual-modality model significantly improved segmentation accuracy, yielding a mean Dice score of 75.6% and an F1 score of 72.6% on the internal test set. Additionally, volumetric analysis showed a high Pearson correlation coefficient (0.94) between predicted and actual lesion volumes. Conclusions: The improved performance of the dual-modality model suggests the benefit of integrating FLAIR MRI to capture lesion characteristics in detecting and segmenting sub-acute and chronic stroke lesions. This could lead to more accurate assessment of brain damage and more effective rehabilitation plans for stroke patients. Future research should focus on larger multi-modal datasets and further investigate segmentation challenges, as well as clinical validation.