Brain development, from embryogenesis to adulthood, is intricate and pivotal for understanding neurodevelopmental disorders and cognitive function. Magnetic resonance imaging (MRI) is essential because of its ability to provide high-detail images without the need for invasive procedures. However, MRI images may degrade, requiring enhancement through supervised techniques. These methods aim to sharpen and upscale images affected by factors such as equipment limitations and sensor quality. Traditionally, super-resolution reconstruction of MRI images has relied on three-dimensional (3D) convolutional neural networks (CNNs). Nonetheless, these networks pose challenges due to their extensive parameters and weight, leading to high memory consumption. Addressing this concern, our study introduces an innovative approach employing two 2D-CNNs. A super-resolution image reconstruction technique is presented that evaluates both conventional and deep learning–based methods, including bicubic interpolation, enhanced deep super-resolution (EDSR), super-resolution generative adversarial network (SRGAN), enhanced super-resolution generative adversarial network (ESRGAN), and receptive field block ESRGAN (RFB-ESRGAN), applied to the T1 brain image dataset. In supervised image reconstruction, prior knowledge about image degradation is harnessed to enhance sharpness and resolution. Notably, our method works with 2D slices and employs a two-step reconstruction process to assemble them into a 3D representation. This approach offers several advantages, including reduced data dimensionality, simplified data handling, and potentially lower computational costs compared to direct 3D data processing. Results indicate that RFB-ESRGAN outperforms other methods, achieving an average peak signal-to-noise ratio (PSNR) of 34.5 across all three image planes and an average structural similarity index (SSIM) of 0.89. Based on these findings, RFB-ESRGAN is selected as the initial step in the reconstruction pipeline. Further experimentation with multiple configurations shows that noise-augmented ESRGAN (nESRGAN), using a depth of 16 and incorporating both noise and interpolation sampling, delivers superior performance. This configuration attains an average PSNR of 34.7 across all image planes and an average SSIM value of 0.91, along with favorable learned perceptual image patch similarity (LPIPS) values. The resulting high-resolution 3D brain reconstructions exhibit enhanced structural fidelity and demonstrate strong potential for surgical applications by providing improved visualization of anatomical details essential for preoperative planning and intraoperative navigation.
Sikkandar et al. (Thu,) studied this question.