The dataset comprises a total of 500 patients, with 300 patients used for model training and 200 patients reserved for model testing. Each patient had approximately 350 2D brain slices of the coregistered 3D volumes excluding the five slices at the base and five slices at the top of the acquired volume due to their low signal-to-noise ratio (SNR). For robustness and to avoid overfitting, standard DL data augmentation techniques consisting of eight types of transformations were applied to each case in the training set.

Using the pre-contrast and low-dose Contrast-Enhanced Magnetic Resonance (CE-MRI) images as input and the true full-dose CE-MRI images as the ground truth. the Cycle GAN deep network (DeepGad), was trained to reconstruct the full-dose CE-MRI images from low-dose CE-MRI images.

Image quality using a 5-point Likert scale covering 1 (none), 2 (poor), 3 (moderate), 4 (good), and 5(excellent) ratings

(0) no normal vessels observed, (1) significant decrease in conspicuity with potential impact on diagnosis, (2) mild decrease in conspicuity with unlikely impact on diagnosis, (3) normal conspicuity, (4) mild increase in conspicuity with unlikely impact on diagnosis, and (5) significant increase in conspicuity with potential impact on diagnosis.

(a) no enhancement, (b) homogeneous enhancement, (c) heterogeneous enhancement, (d) ring enhancement, (e) linear enhancement, and (f) other enhancement.