Assessment of Stroke Onset Time Based on DINOv2 Visual Features

Background

Currently, the thrombolytic treatment for acute ischemic stroke (AIS) strictly depends on the time since stroke onset (TSS) being less than 4.5 hours. However, some patients are excluded from thrombolytic treatment due to uncertain TSS. Clinically, diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) mismatch are commonly used to roughly determine TSS.

Methods

In this paper, we propose a method based on DINOv2 to classify the TSS as less than or more than 4.5 hours. We conducted model training and external testing using case data from two hospitals. These hospitals respectively included 226 and 85 cases of TSS less than 4.5 hours, along with an equal number of cases with TSS greater than 4.5 hours. Firstly, we utilized DINOv2 for automatic segmentation of lesions and extraction of visual features from DWI and FLAIR images. Then, the visual features of the lesion area were input into four different machine learning models. Finally, a conclusion on whether the patient's onset time is more or less than 4.5 hours is reached through a weighted voting method.

Results

The results from the external test set show that in lesion segmentation from DWI and FLAIR images, the Dice coefficients were as high as 0.872 and 0.823, respectively. In the judgment of TTS less than 4.5 hours, our approach achieved an accuracy of 0.865, sensitivity of 0.843, and specificity of 0.902.

Conclusion

The assessment of TTS based on DINOv2 visual features demonstrates excellent performance. The results of this approach significantly surpass those of human doctors using the DWI-FLAIR mismatch method. Moreover, it achieves a fully automated process for rapid and efficient handling. This approach is expected to play a key role in treatment decision-making for patients with unknown TSS.