Abstract
Cerebral blood volume (CBV) maps play an important role in the differential diagnosis between glioblastoma recurrence and treatment response. However, it needs high injective velocity. This study aimed to synthesize CBV maps from arterial spin labeling (ASL) and standard MRI sequences using a deep learning method and to validate its discriminating value. A total of 744 MRI scans from 364 patients were included in this retrospective, single-institution study. A three-dimensional (3D) incrementable encoder-decoder network (IEDN) designed for an asymmetrical sample size was trained to synthesize the CBV maps from ASL and the standard MRIs. The synthetic performance was evaluated quantitatively using the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR) and qualitatively using a four-point Likert scale (from 0 to 3). In 96 patients suspected of glioblastoma recurrence vs. treatment response as the external test set from a hospital-based cohort, the difference in the additive value between synthetic CBV maps and ASL to standard MRIs was examined using the Z test. The best algorithm appeared to be achieved by the 3D IEDN with ASL + T1-weighted imaging (T1WI) + T2-weighted imaging (T2WI) + apparent diffusion coefficient (ADC) maps + post-contrast T1WI (SSIM = 88.69 ± 3.97%, PSNR = 32.76 ± 3.39 dB). For the image quality scores, the mean image quality score for all synthetic CBV maps was 2.90. Standard MRI plus synthetic CBV maps had better performance than standard MRI and ASL scans in the differential diagnosis between tumor recurrence and treatment response (p = 0.019). Therefore, 3D IEDN produced qualified synthetic CBV maps without the need for high injective velocity from ASL and standard MRIs. The synthetic CBV maps achieved better performance in the differential diagnosis between glioblastoma recurrence and treatment response.