Feasibility of U-Net model for cerebral arteries segmentation with low-dose computed tomography angiographic images with pre-processing methods.

Subtraction computed tomography angiography (sCTA) can effectively separate enhanced cerebral arteries from similar signal intensity and proximity (i.e., vertebrae and skull). However, sCTA is not considered mainstream because of the high radiation dose generated by the two-scan protocol. We aimed to solve the overexposure problem by training a U-Net-based CA segmentation model using a low-dose computed tomographic angiography (CTA) image-based dataset with various pre-processing methods to achieve a performance similar to that of sCTA. We optimized a non-local means (NLM) algorithm using the coefficient of variation and contrast-to-noise ratio. In addition, datasets were constructed by predicting the CA mask using a semiautomatic thresholding technique based on region growing method. Then, CTA images of 35 (2052 slices), 4 (248 slices), and 5 patients (594 slices) were used, respectively, for the train, validation, and test sets. To evaluate the performance of the U-Net-based CA segmentation model quantitatively according to the constructed dataset, the average precision (AP), intersection over union (IoU), and F1-score were calculated. For the dataset to which both the optimized NLM algorithm and semiautomatic thresholding technique were applied, the segmentation model showed the most improved performance. In particular, the quantitative evaluation of the low-dose CTA image with the NLM algorithm and the semiautomatic thresholding-based U-Net model calculated AP, IoU, and F1-scores of approximately 0.880, 0.955, and 0.809, respectively, which were most similar to the CA segmentation performance of the sCTA technique. The proposed U-Net model provided CA segmentation results without additional radiation exposure. In addition, the selection and optimization of an appropriate pre-processing methods were identified as essential for achieving higher segmentation performance for the U-Net model.