Abstract
Stroke is the second leading cause of death, accounting for 11% of deaths worldwide. Comparing diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) images is important for stroke diagnosis, but most studies have focused on lesion segmentation using DWI. In this study, we compared the performance of lesion segmentation using DWI and ADC images. This study was conducted using a retrospective design A dataset was constructed using data from 360 patients with ischemic stroke collected from Gachon University Gil Medical Center. Artificial intelligence models, U-Net, and a fully connected network (FCN), were used to train each type of image data. The performance of the models was validated using five-fold cross-validation and evaluated based on metrics such as the dice similarity coefficient (DSC), accuracy, precision, and recall. As a result, the U-Net model demonstrated a DSC of 92.13 ± 0.91% on DWI and 83.68 ± 10% on ADC, whereas the FCN model exhibited a DSC of 82.86 ± 1.56% on DWI and 79.26 ± 1.19% on ADC. These metrics indicated that the trained models were suitable for lesion segmentation. A comparative analysis of DWI and ADC based on the trained models revealed similar results across the models, suggesting that lesion segmentation on ADC images is appropriate. For future research, the accuracy of ADC images is recommended to be imporved by utilizing images with different b-values, or training models with datasets that combe DWI and ADC images based on enhanced data.