Automated delineation of acute ischemic stroke lesions on non-contrast CT using 3D deep learning: A promising step towards efficient diagnosis and treatment

We adopt the existing Deep learning architecture to support diagnosing acute ischemic stroke by automatically detecting lesion location on 3D non-contrast CT brain scans. We also investigate the feasibility of the model's applications in the clinical scenario by data analysis. We retrospectively collected 3D non-contrast CT scans of 317 patients with acute ischemic stroke from the China Medical University Hospital. All patients underwent standard baseline non-contrast CT scanning followed by diffusion-weighted imaging. We utilized these data for training the existing model – SwinUNETR, which includes a self-attention module as an encoder and a convolutional-based decoder. Moreover, the software innovatively incorporates uncertainty quantification to enhance model performance. In the test set, the AI model predicted lesion volume with a mean Dice score of 46.7 % compared to diffusion-weighted imaging verified by experts. The model completed the analysis on a 3D non-contrast CT scan in approximately 30 s. The average difference between the model-segmented acute ischemic stroke lesion volume (67.11 ml) and diffusion-weighted imaging lesion volume (35.2 ml) was 27.09 ml. Pearson correlation of lesion volume between prediction and ground truth is 83.46 %. We also found our model has superior performance in the CT scan with lesion volume > 40 ml and 3 h < onset-to-CT time <24 h. Moreover, our approach was applied to the AISD public dataset, yielding a Dice score of 0.619 upon testing. This model could help facilitate timely and accurate diagnosis of acute ischemic stroke in a clinical emergency setting and low-resourced hospital.