Deep Learning Segmentation of Diffusion-Weighted MRI Acute Ischaemic Stroke: A Pragmatic Evaluation Across Three Datasets

Automated segmentation of acute ischaemic stroke (AIS) lesions is critical for clinical decision-making, yet current deep learning approaches often rely on complex preprocessing pipelines—such as skull-stripping (brain extraction) and the integration of multiple MRI sequences (e.g., ADC, FLAIR). This study evaluates whether a pragmatic, simplified deep learning approach can achieve high-accuracy segmentation using only raw DWI data, thereby reducing computational complexity and improving clinical feasibility.

The researchers trained self-configuring nnU-Net models on a large, diverse dataset of 1,744 AIS cases collected from local, national, and open-access sources. They tested four experimental configurations (varying the use of brain extraction and the inclusion of ADC maps) and compared two model architectures: a baseline nnU-Net and a residual encoder (ResEnc) variant. Performance was benchmarked against the DeepISLES ensemble model, a state-of-the-art tool from the 2022 ISLES challenge, using a test set of 436 cases.

The study found that the baseline nnU-Net, trained solely on raw DWI without any preprocessing, achieved a median Dice similarity coefficient (DSC) of 0.84. While the more complex ResEnc architecture provided marginal improvements in DSC, it required significantly more computational resources and longer inference times. Notably, the baseline nnU-Net significantly outperformed the DeepISLES ensemble model, particularly in cases with smaller stroke volumes, and demonstrated much faster inference (approximately 5 seconds per case compared to 235 seconds for DeepISLES). The results suggest that skull-stripping is unnecessary for DWI-based stroke segmentation and may introduce avoidable errors.

By demonstrating that high-accuracy segmentation can be achieved with minimal preprocessing and only a single input sequence, this research provides a streamlined pathway for integrating automated stroke analysis into routine clinical radiology workflows. This approach reduces the technical barriers to deployment, such as the need for co-registration or complex image cleaning, and offers a robust tool for rapid, automated stroke assessment.