Acute stroke lesion segmentation and prediction tasks are of great clinical interest as they can help doctors make better informed time-critical treatment decisions. Automatic segmentation of these lesions is a complex task due to their heterogeneous appearance, dynamic evolution and inter-patient differences. Typically, acute stroke lesion tasks are approached with methods developed for chronic stroke or other brain lesions. However, the pathophysiology and anatomy of acute stroke establishes an inherently different problem that needs special consideration. In this work, we propose a novel deep learning architecture specially designed for acute stroke tasks that involve approximating complex non-linear functions with reduced data. Within our strategy, class imbalance is tackled using a hybrid strategy based on state-of-the-art train sampling strategies designed for other brain lesion related tasks, which is more suited to the anatomy and pathophysiology of acute stroke lesions. The proposed method is evaluated on three unrelated public international challenge datasets (ISLES) without any dataset specific hyper-parameter tuning. These involve the tasks of sub-acute stroke lesion segmentation, acute stroke penumbra estimation and chronic extent prediction from acute MR images. The performance of the proposed architecture is analysed both against similar deep learning architectures from chronic stroke and related biomedical tasks and also by submitting the segmented test images for blind online evaluation on each of the challenges. When compared with the rest of submitted strategies, our method achieves top-rank performance among the best submitted entries in all the three challenges, showing its capability to deal with different unrelated tasks without hyper-parameter tuning. In order to promote the reproducibility of our results, a public version of the proposed method has been released.