<scp>Nonlocal</scp> convolutional block attention module <scp>VNet</scp> for gliomas automatic segmentation

Fang, Ying; Huang, He; Yang, W.; Xu, Xiaomei; Jiang, Weiwei; Lai, Xiaobo

doi:10.1002/ima.22639

Cited by 23 publications

(15 citation statements)

References 47 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Finally, we evaluate the AEMA‐Net model on the BraTS 2020 validation dataset, and then compare it with eight representative brain tumor segmentation models including Variational‐Autoencoder Regularized 3D MultiResUNet (Tang et al), ME‐Net (Zhang et al), NLCA‐VNet (Fang et al), TransBTS (Wang et al) etc 37,63‐69 . Among these models, Guan et al, 67 Fang et al, 68 Huang et al 69 and Wang et al 65 introduce SE, non‐local, CBAM and self‐attention modules to construct their deep segmentation networks, while the remaining four works are non‐attention models. As shown in Table 5, AEMA‐Net achieves the optimal DSC values of 0.896 and 0.839 on whole tumor and core tumor segmentation, and it ranks third on whole enhancing tumor segmentation results.…”

Section: Resultsmentioning

confidence: 99%

“…achieves the optimal DSC accuracy on enhancing tumor segmentation. Meanwhile, the results of AEMA-Net are slightly lower than those of Isensee et al, 55 Brügger et al 38 37,[63][64][65][66][67][68][69] Among these models, Guan et al, 67 Fang et al, 68 Huang et al 69 Additionally, as shown in Table 1-5, all deep neural models have the optimal DSC values on whole tumor segmentation while achieving the lowest results on the enhancing tumor segmentation. Because the whole tumor region is larger than regions of the core and enhancing tumors, it is easier to segment whole tumors for U-Net models.…”

Section: Comparison With Counterparts and State-of-the-art Methodsmentioning

confidence: 96%

See 1 more Smart Citation

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

Zhang

Jiang

et al. 2021

NMR in Biomedicine

View full text Add to dashboard Cite

Automatic brain tumor segmentation on MRI is a prerequisite to provide a quantitative and intuitive assistance for clinical diagnosis and treatment. Meanwhile, 3D deep neural network related brain tumor segmentation models have demonstrated considerable accuracy improvement over corresponding 2D methodologies. However, 3D brain tumor segmentation models generally suffer from high computation cost. Motivated by a recently proposed 3D dilated multi‐fiber network (DMF‐Net) architecture that pays more attention to reduction of computation cost, we present in this work a novel encoder‐decoder neural network, ie a 3D asymmetric expectation‐maximization attention network (AEMA‐Net), to automatically segment brain tumors. We modify DMF‐Net by introducing an asymmetric convolution block into a multi‐fiber unit and a dilated multi‐fiber unit to capture more powerful deep features for the brain tumor segmentation. In addition, AEMA‐Net further incorporates an expectation‐maximization attention (EMA) module into the DMF‐Net by embedding the EMA block in the third stage of skip connection, which focuses on capturing the long‐range dependence of context. We extensively evaluate AEMA‐Net on three MRI brain tumor segmentation benchmarks of BraTS 2018, 2019 and 2020 datasets. Experimental results demonstrate that AEMA‐Net outperforms both 3D U‐Net and DMF‐Net, and it achieves competitive performance compared with the state‐of‐the‐art brain tumor segmentation methods.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Comparison With Counterparts and State-of-the-art Methodsmentioning

confidence: 96%

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

Zhang

Jiang

et al. 2021

NMR in Biomedicine

View full text Add to dashboard Cite

show abstract

“…, and thus the L RFL in Equation (7) can be viewed as a positive one L pRFL . the whole L RFL loss can be calculated as:…”

Section: Negative Region Term Of Rflmentioning

confidence: 99%

“…Feng et al 5 and Isensee et al 6 optimize the UNET model through different hyper-parameter selection strategy, which can reduce the random errors caused by manual setting of hyper-parameters. Many researchers 7,8 improve VNET 9 to segment brain gliomas, which use residual blocks at each stage and replace pooling with convolutions to improve the segmentation result. However, most of the current works focus on improving the network structure, and few of them on how to make the segmentation network focalize hard voxels to improve their performance.…”

Section: Introductionmentioning

confidence: 99%

Region‐related focal loss for 3D brain tumor MRI segmentation

You

Peng

et al. 2023

Medical Physics

View full text Add to dashboard Cite

Background In the brain tumor magnetic resonance image (MRI) segmentation, although the 3D convolution networks (CNNs) has achieved state‐of‐the‐art results, the class and hard‐voxel imbalances in the 3D images have not been well addressed. Voxel independent losses are dependent on the setting of class weights for the class imbalance issue, and are hard to assign each class equally. Region‐related losses cannot correctly focus on hard voxels dynamically and not be robust to misclassification of small structures. Meanwhile, repeatedly training on the additional hard samples augmented by existing methods would bring more class imbalance, overfitting and incorrect knowledge learning to the model. Purpose A novel region‐related loss with balanced dynamic weighting while alleviating the sensitivity to small structures is necessary. In addition, we need to increase the diversity of hard samples in the training to improve the performance of model. Methods The proposed Region‐related Focal Loss (RFL) reshapes standard Dice Loss (DL) by up‐weighting the loss assigned to hard‐classified voxels. Compared to DL, RFL adaptively modulate its gradient with an invariant focalized point that voxels with lower‐confidence than it would achieve a larger gradient, and higher‐confidence voxels would get a smaller gradient. Meanwhile, RFL can adjust the parameters to set where and how much the network is focused. In addition, an Intra‐classly Transformed Augmentation network (ITA‐NET) is proposed to increase the diversity of hard samples, in which the 3D registration network and intra‐class transfer layer are used to transform the shape and intensity respectively. A selective hard sample mining(SHSM) strategy is used to train the ITA‐NET for avoiding excessive class imbalance. Source code (in Tensorflow) is available at: https://github.com/lb-whu/RFL_ITA. Results The experiments are carried out on public data set: Brain Tumor Segmentation Challenge 2020 (BratS2020). Experiments with BraTS2020 online validation set show that proposed methods achieve an average Dice scores of 0.905, 0.821, and 0.806 for whole tumor (WT), tumor core (TC) and enhancing tumor (ET), respectively. Compared with DL (baseline), the proposed RFL significantly improves the Dice scores by an average of 1%, and for the small region ET it can even increase by 3%. And the proposed method combined with ITA‐NET improves the Dice scores of ET and TC by 5% and 3% respectively. Conclusions The proposed RFL can converge with a invariant focalized point in the training of segmentation network, thus effectively alleviating the hard‐voxel imbalance in brain tumor MRI segmentation. The negative region term of RFL can effectively reduce the sensitivity of the segmentation model to the misclassification of small structures. The proposed ITA‐NET can increase the diversity of hard samples by transforming their shape and transfer their intra‐class intensity, thereby effectively improving the robustness of the segmentation network to hard samples.

show abstract

“…40 Several research works have been performed efficiently on 41 the segmentation in B-mode echocardiography in the past 42 few decades [2]- [4]. With the combination of various feature 43 enhancement modules [5], [6] and different deep network 44 architectures [7]- [10], the ground-truth is applied as a class 45 associate or shape regulation by minimizing the loss function. 46 However, these methods still have scope for improvement.…”

Section: Introductionmentioning

confidence: 99%

MCAL: An Anatomical Knowledge Learning Model for Myocardial Segmentation in 2-D Echocardiography

Cui

Zhang

et al. 2022

IEEE Trans. Ultrason., Ferroelect., Freq. Contr.

View full text Add to dashboard Cite

Segmentation of the left ventricular (LV) my-1 ocardium in 2D echocardiography is essential for clinical decision 2 making, especially in geometry measurement and index computa-3 tion. However, segmenting the myocardium is a time-consuming 4 process as well as challenging due to the fuzzy boundary 5 caused by the low image quality. Previous methods based on 6 deep Convolutional Neural Networks (CNN) employ the ground-7 truth label as class associations on the pixel-level segmentation, 8 or use label information to regulate the shape of predicted 9 outputs, works limit for effective feature enhancement for 2D 10 echocardiography. We propose a training strategy named multi-11 constrained aggregate learning (referred as MCAL), which lever-12 ages anatomical knowledge learned through ground-truth labels 13 to infer segmented parts and discriminate boundary pixels. The 14 new framework encourages the model to focus on the features in 15 accordance with the learned anatomical representations, and the 16 training objectives incorporate a Boundary Distance Transform 17 Weight (BDTW) to enforce a higher weight value on the boundary 18 region, which helps to improve the segmentation accuracy. The 19 proposed method is built as an end-to-end framework with a 20 top-down, bottom-up architecture with skip convolution fusion 21 blocks, and carried out on two datasets (our dataset and the 22 public CAMUS dataset). The comparison study shows that the 23 proposed network outperforms the other segmentation baseline 24 models, indicating that our method is beneficial for boundary 25 pixels discrimination in segmentation. 26 Index Terms-Boundary distance transform weight, multi-27 constrained aggregate learning, myocardial segmentation. 28 encoder. Furtherly adjusting the scale and offset of the learned 68 anatomical information by a Feature-wise Linear Modulation 69 (FiLM) [11], the segmented relevance feature information 70 is enhanced. Finally, upon observing that the boundary is 71 hard to detect, we further enhance a higher weight on the 72 border neighborhood pixels by proposing a Boundary Distance 73 Transform Weight (BDTW) for the segmentation loss, which 74 acts as guidance for penalizing the learning process. 75 The main contributions of our proposed framework are: 76 • Our method derives segmented-relevance information 77 by narrowing distribution divergence between the latent 78 space of input and label, which exploits anatomical 79 knowledge to guide feature enhancement. FiLM is ap-80 plied to enhance segmented-relevant features with the 81 guidance of the anatomical information, which restrains 82 the irrelative features under low image quality. 83 • A novel Boundary Distance Transform Weight (BDTW) 84 is applied to the cross-entropy loss. It forces the network 85 to focus on the boundary region pixels in each training 86 batch and improves the discrimination on boundary pix-87 els, which is useful in cases with low image quality. 88 II. RELATED WORKS 89 There have been many works on the segmentation of B-90 mode echoc...

show abstract

Nonlocal convolutional block attention module VNet for gliomas automatic segmentation

Cited by 23 publications

References 47 publications

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

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

Region‐related focal loss for 3D brain tumor MRI segmentation

MCAL: An Anatomical Knowledge Learning Model for Myocardial Segmentation in 2-D Echocardiography

Contact Info

Product

Resources

About