3D asymmetric expectation‐maximization attention network for brain tumor segmentation

Zhang, Jianxin; Jiang, Zongkang; Li, Dongwei; Sun, Qiule; Hou, Yaqing; Liu, Bin

doi:10.1002/nbm.4657

Cited by 10 publications

(4 citation statements)

References 69 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Considering the limitations of 2D U-Net in understanding spatial context features of 3D MRI brain tumor images, researchers proposed a 3D U-Net 21 and greatly improved the segmentation performance for brain tumors. In addition, researchers have introduced autoencoders, 22 attention mechanism, [23][24][25] and cascade architecture [26][27][28][29][30][31] into networks to solve insufficient representations of highresolution features in small-scale and irregular brain tumor regions based on 3D U-Net, which greatly promoted the development of 3D U-Net for BraTS. Among them, Myronenko 26 combined variational autoencoder (VAE) with 3D U-Net, effectively improving the accuracy of BraTS, and this method won first place in the 2018 BraTS Challenge.…”

Section: Related Workmentioning

confidence: 99%

An efficient brain tumor segmentation model based on group normalization and 3D U‐Net

Chen,

Lin,

Ren

et al. 2024

Int J Imaging Syst Tech

View full text Add to dashboard Cite

Accurate segmentation of brain tumors has a vital impact on clinical diagnosis and treatment, and good segmentation results are helpful for the treatment of this disease, which is a serious threat to human health. High‐precision segmentation of brain tumors remains a challenging task due to their diverse shapes, sizes, locations, and complex boundaries. Considering the special structure of medical brain tumor images, many researchers have proposed a brain tumor segmentation (BraTS) network based on 3D U‐Net. However, there are also problems such as insufficient receptive fields and excessive computing costs. In this paper, we propose an efficient BraTS model based on group normalization (GN) and 3D U‐Net (3D‐EffUNet). First, according to the characteristics of brain tumor images, the medical image of the whole case is input into the model, and 3D convolution layers are used to extract features and filter irrelevant information. Then, using 3D U‐Net as the main framework, an efficient convolutional module is designed for more precise processing of brain tumor features. Moreover, an efficient convolution module based on GN and an attention mechanism is introduced to reduce the complexity of the network without affecting the segmentation performance and to increase the awareness of voxels between adjacent dimensions and the local space. Finally, the decoder was used to reconstruct high‐precision BraTS information. The model is trained and tested on the BraTS2021 dataset, and the experimental results show that it can maintain good segmentation performance and greatly reduce the calculation cost.

show abstract

Section: Related Workmentioning

confidence: 99%

An efficient brain tumor segmentation model based on group normalization and 3D U‐Net

Chen,

Lin,

Ren

et al. 2024

Int J Imaging Syst Tech

View full text Add to dashboard Cite

show abstract

“…First, the image appearance of 6-month-old infant brain MR images is quite noisy (Li et al, 2019;Mostapha & Styner, 2019) which makes the effective feature extraction difficult for the traditional convolution kernel design in previous works. Adopting enhanced convolution kernel designs (Ding et al, 2019;Li et al, 2020;Zhang et al, 2022) the anatomical morphology of gyrus at large spatial scales (Wang, Li, Adeli, et al, 2018). Although previous infant segmentation approaches try to fuse multi-scale features by skip-connections in variants of FCN and U-Net, they overlook capturing rich multi-scale features in kernel space, which contains more stable and homogeneous semantic information than features between layers (Fan et al, 2019).…”

Section: Improvements From Fine-grained Convolution Kernel Designs Ar...mentioning

confidence: 99%

“…First, the image appearance of 6‐month‐old infant brain MR images is quite noisy (Li et al, 2019; Mostapha & Styner, 2019) which makes the effective feature extraction difficult for the traditional convolution kernel design in previous works. Adopting enhanced convolution kernel designs (Ding et al, 2019; Li et al, 2020; Zhang et al, 2022) that emphasizes key features in the skeleton center of kernels may facilitate feature extractions throughout the network. Second, the voxel‐wise fuzzy tissue boundaries in infant brain images are constrained by the anatomical morphology of gyrus at large spatial scales (Wang, Li, Adeli, et al, 2018).…”

Section: Introductionmentioning

confidence: 99%

3D‐MASNet: 3D mixed‐scale asymmetric convolutional segmentation network for 6‐month‐old infant brain MR images

Zeng

Zhao

Sun

et al. 2022

Human Brain Mapping

View full text Add to dashboard Cite

Precise segmentation of infant brain magnetic resonance (MR) images into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) are essential for studying neuroanatomical hallmarks of early brain development. However, for 6‐month‐old infants, the extremely low‐intensity contrast caused by inherent myelination hinders accurate tissue segmentation. Existing convolutional neural networks (CNNs) based segmentation models for this task generally employ single‐scale symmetric convolutions, which are inefficient for encoding the isointense tissue boundaries in baby brain images. Here, we propose a 3D mixed‐scale asymmetric convolutional segmentation network (3D‐MASNet) framework for brain MR images of 6‐month‐old infants. We replaced the traditional convolutional layer of an existing to‐be‐trained network with a 3D mixed‐scale convolution block consisting of asymmetric kernels (MixACB) during the training phase and then equivalently converted it into the original network. Five canonical CNN segmentation models were evaluated using both T1‐ and T2‐weighted images of 23 6‐month‐old infants from iSeg‐2019 datasets, which contained manual labels as ground truth. MixACB significantly enhanced the average accuracy of all five models and obtained the most considerable improvement in the fully convolutional network model (CC‐3D‐FCN) and the highest performance in the Dense U‐Net model. This approach further obtained Dice coefficient accuracies of 0.931, 0.912, and 0.961 in GM, WM, and CSF, respectively, ranking first among 30 teams on the validation dataset of the iSeg‐2019 Grand Challenge. Thus, the proposed 3D‐MASNet can improve the accuracy of existing CNNs‐based segmentation models as a plug‐and‐play solution that offers a promising technique for future infant brain MRI studies.

show abstract

“…First, the image appearance of 6-month-old infant brain MR images is quite noisy (Li et al, 2019; Mostapha and Styner, 2019) which makes the effective feature extraction difficult for the traditional convolution kernel design in previous works. Adopting enhanced convolution kernel designs (Ding et al, 2019; Li et al, 2020; Zhang et al, 2022) that emphasizes key features in the skeleton center of kernels may facilitate feature extractions throughout the network. Second, the voxel-wise fuzzy tissue boundaries in infant brain images are constrained by the anatomical morphology of gyrus at large spatial scales (Wang et al, 2018a).…”

Section: Introductionmentioning

confidence: 99%

3D-MASNet: 3D Mixed-scale Asymmetric Convolutional Segmentation Network for 6-month-old Infant Brain MR Images

Zeng

Zhao

Sun

et al. 2021

Preprint

View full text Add to dashboard Cite

Precise segmentation of infant brain MR images into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) is essential for studying neuroanatomical hallmarks of early brain development. However, for 6-month-old infants, the extremely low-intensity contrast caused by inherent myelination hinders accurate tissue segmentation. Existing convolutional neural networks (CNNs) based segmentation model for this task generally employ single-scale symmetric convolutions, which are inefficient for encoding the isointense tissue boundaries in limited samples of baby brain images. Here, we propose a 3D mixed-scale asymmetric convolutional segmentation network (3D-MASNet) framework for brain MR images of 6-month-old infant. We replaced the traditional convolutional layer of an existing to-be-trained network with a 3D mixed-scale convolution block consisting of asymmetric kernels (MixACB) during the training phase and then equivalently converted it into the original network. Five canonical CNN segmentation models were evaluated using both T1- and T2-weighted images of 23 6-month-old infants from iSeg-2019 datasets, which contained manual labels as ground truth. MixACB significantly enhanced the average accuracy of all five models and obtained the largest improvement in the fully convolutional network model (CC-3D-FCN) and the highest performance in the Dense U-Net model. This approach further obtained Dice coefficient accuracies of 0.931, 0.912, and 0.961 in GM, WM, and CSF, respectively, ranking first among 30 teams on the validation dataset of the iSeg-2019 Grand Challenge. Thus, the proposed 3D-MASNet can improve the accuracy of existing CNNs-based segmentation models as a plug-and-play solution that offers a promising technique for future infant brain MRI studies.

show abstract

3D asymmetric expectation‐maximization attention network for brain tumor segmentation

Cited by 10 publications

References 69 publications

An efficient brain tumor segmentation model based on group normalization and 3D U‐Net

An efficient brain tumor segmentation model based on group normalization and 3D U‐Net

3D‐MASNet: 3D mixed‐scale asymmetric convolutional segmentation network for 6‐month‐old infant brain MR images

3D-MASNet: 3D Mixed-scale Asymmetric Convolutional Segmentation Network for 6-month-old Infant Brain MR Images

Contact Info

Product

Resources

About