Source-linked AI summary
Attention-Guided Version of 2D UNet for Automatic Brain Tumor Segmentation
Mehrdad Noori, Ali Bahri, Karim Mohammadi
TL;DR
Manual MRI tumor segmentation is time-consuming and variable, and 2D models omit 3D context while naive feature fusion may confuse the network. The paper adds channel attention and Multi-View Fusion to a low-parameter 2D UNet. It reports favorable performance against 2017 and 2018 state-of-the-art methods.
Problem
Manual 3D MRI tumor segmentation is time-consuming and variable, while 2D models do not benefit from 3D context and naive feature integration may confuse the model.
Source: §I · pp. 1–2Method
The paper extends a low-parameter 2D UNet with channel attention after feature concatenation and Multi-View Fusion of axial and coronal views.
Source: §I · p. 2, C. Multi-view Fusion · p. 3Results
The method performs favorably against 2017 and 2018 state-of-the-art methods, with a BRATS 2018 validation average ET Dice score of 0.813.
Source: §I · p. 2, §IV · p. 6Takeaways & Limitations
The proposed combination lets a 2D architecture benefit from 3D contextual information while weighting concatenated features adaptively.
Source: C. Multi-view Fusion · p. 3, §IV · p. 6
Abstract
from arXiv · showhide
Gliomas are the most common and aggressive among brain tumors, which cause a short life expectancy in their highest grade. Therefore, treatment assessment is a key stage to enhance the quality of the patients' lives. Recently, deep convolutional neural networks (DCNNs) have achieved a remarkable performance in brain tumor segmentation, but this task is still difficult owing to high varying intensity and appearance of gliomas. Most of the existing methods, especially UNet-based networks, integrate low-level and high-level features in a naive way, which may result in confusion for the model. Moreover, most approaches employ 3D architectures to benefit from 3D contextual information of input images. These architectures contain more parameters and computational complexity than 2D architectures. On the other hand, using 2D models causes not to benefit from 3D contextual information of input images. In order to address the mentioned issues, we design a low-parameter network based on 2D UNet in which we employ two techniques. The first technique is an attention mechanism, which is adopted after concatenation of low-level and high-level features. This technique prevents confusion for the model by weighting each of the channels adaptively. The second technique is the Multi-View Fusion. By adopting this technique, we can benefit from 3D contextual information of input images despite using a 2D model. Experimental results demonstrate that our method performs favorably against 2017 and 2018 state-of-the-art methods.
I. INTRODUCTION
Manual brain-tumor MRI segmentation is time-consuming and variable, while 2D models lack 3D context and naive feature fusion may confuse the network. The paper addresses these gaps with Multi-View Fusion and channel attention in a 2D UNet, reporting favorable comparisons with 2017 and 2018 state-of-the-art methods.
I. INTRODUCTION
- Manual segmentation of 3D MRI scans is time-consuming and susceptible to inaccuracies and variability because tumor appearance is complex.Automatic segmentation can inform diagnosis, growth-rate prediction, and treatment planning, particularly where expert radiologist access is restricted.Source: §I · p. 1
- Prior work described here spans 2D multi-path CNNs, 3D multi-scale networks, and encoder-decoder approaches for brain-tumor segmentation.The related-work discussion also covers methods used by BRATS challenge participants.Source: §I · pp. 1–2
- The paper combines a low-parameter 2D UNet with Multi-View Fusion to retain 3D contextual information despite processing 2D slices.The authors motivate the 2D architecture by its lower parameter count and computational complexity than 3D UNet.Source: §I · p. 2
- Channel attention after concatenating multi-level features adaptively weights channels to address potential confusion from treating feature maps as equally important.The authors identify this naive equal weighting as a potential source of confusion.Source: §I · p. 2
- Although the method is 2D, it performs favorably against state-of-the-art methods from both 2017 and 2018.Source: §I · p. 2
A. Network
The network uses a UNet-style contracting and expanding design, with residual units and channel attention to combine encoder and decoder features. Strided convolutions replace pooling for downsampling, while squeeze-and-excitation blocks reweight concatenated channels.
A. Network
- The UNet-based network has a contracting path with three downsampling layers and an expanding path that restores the input resolution.Stride-2 convolutions replace the original UNet’s max-pooling layers.Source: A. Network · p. 2
- Residual units replace plain units, using two convolutional layers followed by batch normalization and PReLU; the bottleneck and expanding path also use residual units.The authors state this choice is intended to speed training and convergence.Source: A. Network · p. 2
- Each expanding-path upsampling layer doubles feature-map size and is followed by a 2×2 convolution.Source: A. Network · p. 3
- After each skip-feature concatenation, a squeeze-and-excitation block adaptively weights channels before the features proceed through the network.The block reweights concatenated high- and low-level feature maps; the final 1×1 convolution and Softmax map features to the class outputs.Source: A. Network · p. 3, Figure 2 · p. 3
B. Loss Function
The loss combines Generalized Dice loss and Cross-entropy to address class imbalance in brain tumor segmentation. Generalized Dice loss adaptively weights classes, while the combined objective uses λ = 1.25.
B. Loss Function
- The training objective combines Generalized Dice loss with Cross-entropy loss to address the task’s class imbalance.The authors state that Cross-entropy speeds convergence.Source: B. Loss Function · p. 3
- λ is empirically set to 1.25 in the combined loss.Source: B. Loss Function · p. 3
- Generalized Dice loss adaptively weights classes to address imbalances in brain tumor classes.It is combined with Cross-entropy, which the authors say speeds convergence.Source: B. Loss Function · p. 3
- Generalized Dice loss is a multi-class version of Dice loss.Source: B. Loss Function · p. 3
C. Multi-view Fusion
The Multi-view Fusion procedure uses separate 2D models for axial and coronal slices, then combines their reconstructed 3D outputs. Pixel-wise averaging lets the 2D approach consider 3D image context.
C. Multi-view Fusion
- Separate 2D networks are trained on axial and coronal MRI slices; their predicted maps are concatenated into 3D volumes and fused by pixel-wise averaging.The authors state this fusion lets the method consider the 3D nature of the input images.Source: C. Multi-view Fusion · p. 3
III. EXPERIMENTS
Experiments use BRATS 2017 and 2018 MRI datasets, with results evaluated for three tumor sub-regions using Dice Score and Hausdorff Distance.
A. Dataset description
- The BRATS 2017 and 2018 training sets contain 285 patients’ MRI volumes, while their validation sets contain 46 and 66 patients, respectively.The training set comprises 210 HGG and 75 LGG cases; validation-set patient grades are unknown.Source: A. Dataset description · p. 3
- Evaluation merges labels into Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) regions, using expert manual segmentations as ground truth.Validation labels are not public; quantitative evaluation is obtained through the BRATS online platform.Source: A. Dataset description · p. 4
B. Evaluation metrics
- Dice Score and Hausdorff Distance evaluate ET, WT, and TC; Hausdorff95 uses the 95th percentile instead of the maximum to reduce problems with noisy predictions.Dice measures similarity between sets, while Hausdorff Distance is defined as the longest distance from a point in one set to its nearest point in the other.Source: B. Evaluation metrics · p. 4
C. Implementation Details
The implementation uses five-fold cross-validation on BRATS 2018, with separate view-specific networks and standard preprocessing, augmentation, and SGD training.
C. Implementation Details
- For each view, five-fold cross-validation trains on 228 of 285 BRATS 2018 cases per fold, and the five networks’ Softmax outputs are averaged for evaluation.The BRATS 2017 and 2018 validation results were computed by the online evaluation platform.Source: C. Implementation Details · p. 4
- Preprocessing applies N4ITK bias-field correction, removes the top and bottom 1% of intensities, and normalizes each modality to zero mean and unit variance.The text continues the preprocessing description across the two passages.Source: C. Implementation Details · p. 4
- Training uses random vertical and horizontal flips, Gaussian input noise with standard deviation 0.01, and SGD with momentum 0.9 and learning rate 8e-3.The network was developed in Keras with a TensorFlow backend.Source: C. Implementation Details · p. 4
D. Ablation Study
The ablation study evaluates minor architectural modifications and channel attention, reporting improved overall performance with attention.
D. Ablation Study
- The attention mechanism improves overall performance significantly in the first-fold, Axial-view BRATS 2018 validation experiment.The passage attributes the improvement to weighting multi-level features.Source: D. Ablation Study · p. 4
- The ablation study groups modifications into Minor Modifications and an attention mechanism that weights multi-level features using SE Blocks.The minor modifications include Residual Units, strided convolution, PReLU, and BN.Source: D. Ablation Study · p. 4
- All models in this analysis were trained on the first fold and Axial view of the BRATS 2018 training set, then evaluated on its validation set.Source: D. Ablation Study · p. 4
E. The Proposed Method Results
Multi-view fusion improves the overall results, especially on Hausdorff distance. The section presents comparisons across Axial and Coronal views and their fusion, including visual examples and score distributions.
E. The Proposed Method Results
- Multi-view Fusion improves overall performance, especially on Hausdorff distance, indicating a benefit from considering the data’s 3D nature.Table II reports 5-fold cross-validation ensemble results for Axial and Coronal views and Multi-view Fusion.Source: E. The Proposed Method Results · p. 4, Table II · p. 4
- Fig. 4 compares Axial, Coronal, and Multi-view Fusion segmentation examples for one HGG tumor and one LGG tumor.The examples show input slices and segmentation masks in Axial and Coronal views.Source: E. The Proposed Method Results · p. 4, Figure 4 · p. 5
- Fig. 5 presents Dice score and Hausdorff95 box plots for the three tumor regions across Axial, Coronal, and Multi-view Fusion results.The plots show the 5-fold cross-validation ensemble results.Source: Figure 5 · p. 5, E. The Proposed Method Results · p. 6
F. Comparison with the Existing Methods
On BRATS 2017 and 2018 validation datasets, the 2D network performs favorably against challenge methods, particularly for ET Dice and Hausdorff95. The authors report BRATS 2018 validation average Dice scores for ET, WT, and TC.
F. Comparison with the Existing Methods
- The 2D network performs favorably against BRATS 2017 and 2018 state-of-the-art methods, especially on ET Dice and Hausdorff95 across all three sub-regions.Evaluation uses the challenges’ validation datasets and compares Dice score and Hausdorff distance.Source: F. Comparison with the Existing Methods · p. 6
IV. CONCLUSION
- The method combines a low-parameter 2D UNet with Multi-View contextual information and attention to select discriminative features after low- and high-level feature concatenation.The attention mechanism is intended to prevent confusion from treating all concatenated feature maps as equally important.Source: §IV · p. 6
- 0.813, 0.895, and 0.823 are the average Dice scores for ET, WT, and TC, respectively, on the BRATS 2018 validation dataset.The conclusion attributes these results to adopting the proposed techniques.Source: §IV · p. 6