Automatic Segmentation of Ischaemic Stroke Lesions Using Transformers and Convolutional Neural Networks Applied to...

Ischaemic stroke constitutes a leading cause of global disability. Rapid extraction of the infarct core from multimodal computed tomography perfusion (CTP) imaging guides reperfusion therapy and clinical decision-making. Deep learning algorithms automate this delineation, yet hospital translation is hindered by high-dimensional data, inter-scanner variability, and the low contrast of early ischaemia. Architectural comparisons in the literature frequently carry methodological biases originating from disparate preprocessing protocols and data partitions. This study reduces these variables by evaluating three segmentation strategies under a shared preprocessing pipeline and an identical data partition using the ISLES 2024 dataset. Three models were trained on the same 133-patient partition using a shared preprocessing pipeline based on morphological skull-stripping and modality-specific clinical intensity ranges. The data, preprocessing, and partitions are held constant across models, while framework-dependent factors (optimiser, patch size, physical field of view, spatial resampling, augmentation policy, and model capacity) remain coupled to each architecture and are therefore treated as part of the compared strategy rather than as fully isolated variables. The first of these is a single-stage 5-channel nnU-Net, followed by a two-stage cascaded nnU-Net (2 and 7 channels) and a lightweight Transformer (SegFormer3D). Evaluation on a fixed 15-patient held-out test set isolated the architectural performance. The cascade model achieved the highest Dice Similarity Coefficient (0.224). The single-stage nnU-Net provided the most precise volumetric estimation, recording an Absolute Volume Difference (AVD) of 23.70 mL and a lesion-wise F1-score of 7.60%. On the other hand, SegFormer3D returned the lowest overall metrics (DSC 0.163, AVD 27.28 mL, F1 2.30%). In the small held-out cohort, paired statistical testing did not reveal significant differences between models, so the reported orderings describe the present dataset and experimental configuration rather than a general architectural law. Within these limits, the local inductive bias of the convolutional models retained an empirical advantage over the single Transformer evaluated when processing this moderately sized neuroimaging dataset, and complex cascade topologies offered only marginal gains compared with a well-calibrated single-stage network. Although the predictive segmentation of infarcted tissue at acute stages still demands computational improvements, these results suggest that preprocessing quality is at least as decisive for clinical impact as increasing the complexity of neural architectures.