Brain Extraction From Brain MRI Images Based on Wasserstein GAN and O-Net

Abstract

Brain extraction is an essential pre-processing step for neuroimaging analysis. It is difficult to achieve high-precision extraction from low-quality brain MRI images with artifacts and gray inconsistencies which often result in irregular hole regions in the extracted brain tissues. In addition, the U-Net based brain extraction methods trend to output over-smoothed brain boundary. To remove those irregular holes in the extracted mask, we proposed a new U-Net based model for brain extraction named O-Net. O-Net replaces the skip-connection path in the U-Net with a dual shortcut paths including an attention module to form an Oshaped network, which uses deep semantic information to highlight the target area while retaining more image details. O-Net effectively reduces the impact of intensity differences resulted from artifacts or gray inconsistencies in the brain MRI images on the extraction results. To obtain more precise brain boundary, we designed a new GAN based brain extraction method, which used above O-Net as the segmentation network. The discriminant network of the proposed GAN model adopts the residual structure to enhance the nonlinear expression ability of the network to balance the adversarial training of the two networks. To speed up the convergence of the proposed model, we added a segmentation loss to the adversarial loss to supervise the feature learning of the segmentation network. This method was compared with other popular brain extraction methods on two public datasets (IBSR18 and LPBA40). The mean dice similarity coefficients obtained by the proposed method were 97.26% and 98.29% on IBSR18 and LPBA40 respectively. The results of the proposed method are the best on the two public datasets. Experimental results prove that the proposed model can stably output high-precision brain tissue and reduce the influence of artifacts and gray inconsistencies.

Dimensionality Reduction Hybrid U-Net for Brain Extraction in Magnetic Resonance Imaging

“…Figure 7 shows a set of raw slices with benchmark labels, and we can see that there is a good match between the brain and label. This dataset has been generally used for deep learning brain extraction algorithms [24,26,34], which makes it suitable for testing the algorithm in this manuscript. According to the segmentation results of previous studies, we can see good performance with UNet with this dataset.…”

“…In addition, there are currently several directions for optimizing the U-Net using existing modules. These include adding the residual modules to the U-Net network [24,25] to enhance its learning ability, incorporating attention blocks [26] into the encoding and decoding process, and utilizing advanced U-Net connection methods [27] to improve information interaction across different levels. Furthermore, some studies [28] suggest that the U-Net cascade method is highly effective, and optimizing the preprocessing and postprocessing are very important.…”

PCSS: Skull Stripping With Posture Correction From 3D Brain MRI for Diverse Imaging Environment

“…One of the significant challenges in achieving robust SS is dealing with variability of the subject posture, which can differ significantly depending on different imaging environments and disease status between datasets. This variability can lead to geometric differences between the training and test data, making accurate extraction of brain structure more difficult, even with deep learning-based SS methods [23], [24]. Despite this known issue, no SS methodologies have been proposed that adequately account for the diversity of subject postures.…”

“…In this paper, we propose posture correction skull stripping (PCSS), a highly accurate and robust SS method that takes into account the diversity in subject postures. PCSS is a framework based on U-Net with the following four extensions: (i) preprocessing to estimate and correct the angle and position of the subject's head to suppress posture variation; (ii) weighted loss function, which considers the imbalance between the brain region and other tissues [22]; (iii) a discriminator network for adversarial training introduced in generative adversarial networks (GANs) [31] and used in an SS study [24]; and (iv) ensemble of three-way segmentation for the brain [29]. For a rigorous evaluation across multiple datasets, we use five T1-weighted brain MRI public datasets (ADNI, CC-12, LPBA40, NFBS, and OASIS) and discuss the impact of different datasets on SS performance, which has not been addressed previously.…”

Data augmentation for imbalanced blood cell image classification

“…Advanced versions of WGAN, such as Wasserstein GAN with gradient penalty 37 (WGAN-GP) and Wasserstein divergence GAN 38 (WGAN-div), offer more stable optimisation processes and realistic synthetic images. Therefore, recent studies have employed them for various applications such as preprocessing of brain MRI images 39 , image quality improvement of X-ray images 40 and segmentation of fundus images 41 . However, application of WGAN in biomedical studies to handle data imbalance has not been explored much.…”