Determining the penumbra, i.e., the at-risk but salvageable tissue, is crucial in the context of acute ischemic stroke imaging. Deep learning methods performing segmentation from perfusion parameter maps have shown promise in this regard. However, these methods rely on the computation of parameter maps via deconvolution algorithms, raising concerns about their generalizability across different medical centers. This study investigates the robustness of segmentation methods given different perfusion processing algorithms for dynamic susceptibility contrast magnetic resonance perfusion imaging. A neural network is first trained on a dataset of 94 patients with paired Tmax maps from a single MR perfusion algorithm, together with manual perfusion deficit segmentations. The network’s outputs are then compared on a second dataset of 268 patients, where Tmax inputs are generated with three different deconvolution algorithms. DICE coefficient along with the difference between estimated perfusion deficit volumes are used to quantify the agreement between predictions. Our findings demonstrate high variability in the predicted penumbra, even when Tmax inputs exhibit high similarity (SSIM > 0.8). This study therefore highlights the importance of exploring deconvolution-free methods to address the robustness issue for learning-based penumbra segmentation.