PaperlayerView exact revision on arXiv

Source-linked AI summary

3D AGSE-VNet: An Automatic Brain Tumor MRI Data Segmentation Framework

Xi Guan, Guang Yang, Jianming Ye, Weiji Yang, Xiaomei Xu, Weiwei Jiang, Xiaobo Lai

arXiv:2107.12046v1cs.AIcs.CVcs.LG

TL;DR

Manual or prior tumor segmentation is time-consuming, labor-intensive, error-prone, and limited for multi-modal images. The paper presents AGSE-VNet for automatic 3D MRI brain-tumor segmentation, reporting good BraTS 2020 results while noting weaker performance in enhancing-tumor and tumor-core areas.

  • Problem

    Brain-tumor segmentation requires professional knowledge and is time-consuming, labor-intensive, and prone to errors, with shortcomings for multi-modal images.

  • Method

    AGSE-VNet is an automatic 3D MRI brain-tumor segmentation method using feature maps from different resolutions and attention-guided processing.

  • Results

    The method automatically segments three tumor regions and achieved good results on the BraTS 2020 dataset.

  • Takeaways & Limitations

    AGSE-VNet provides an automatic method for segmenting three brain-tumor regions from 3D MRI images.

  • Takeaways & Limitations

    The paper notes shortcomings for multi-modal images and weaker performance in the enhancing-tumor and tumor-core areas.

Abstract

from arXiv · show

Background: Glioma is the most common brain malignant tumor, with a high morbidity rate and a mortality rate of more than three percent, which seriously endangers human health. The main method of acquiring brain tumors in the clinic is MRI. Segmentation of brain tumor regions from multi-modal MRI scan images is helpful for treatment inspection, post-diagnosis monitoring, and effect evaluation of patients. However, the common operation in clinical brain tumor segmentation is still manual segmentation, lead to its time-consuming and large performance difference between different operators, a consistent and accurate automatic segmentation method is urgently needed. Methods: To meet the above challenges, we propose an automatic brain tumor MRI data segmentation framework which is called AGSE-VNet. In our study, the Squeeze and Excite (SE) module is added to each encoder, the Attention Guide Filter (AG) module is added to each decoder, using the channel relationship to automatically enhance the useful information in the channel to suppress the useless information, and use the attention mechanism to guide the edge information and remove the influence of irrelevant information such as noise. Results: We used the BraTS2020 challenge online verification tool to evaluate our approach. The focus of verification is that the Dice scores of the whole tumor (WT), tumor core (TC) and enhanced tumor (ET) are 0.68, 0.85 and 0.70, respectively. Conclusion: Although MRI images have different intensities, AGSE-VNet is not affected by the size of the tumor, and can more accurately extract the features of the three regions, it has achieved impressive results and made outstanding contributions to the clinical diagnosis and treatment of brain tumor patients.

1. Introduction

Glioma MRI segmentation is clinically important but manual annotation is time-consuming, error-prone, and operator-dependent. The paper proposes 3D AGSE-VNet, combining SE and AG modules within VNet for automatic multimodal MRI segmentation.

  • MRI provides high-contrast soft-tissue images, high spatial resolution, and multiple sequence information from one scan.
  • Manual brain-tumor segmentation requires expertise, is labor-intensive and error-prone, and depends strongly on the doctor's experience.
  • VNet performs well on single-modal images but has shortcomings for multimodal segmentation.
  • 3D AGSE-VNet combines VNet with SE and AG modules for automatic brain-tumor MRI data segmentation.
  • The framework uses volumetric input and three-dimensional convolution to process MRI images, with reported potential for clinical application.

2. Related Works

Earlier brain-tumor segmentation methods include thresholding, region-based, boundary-based, machine-learning, and deep-learning approaches, each with reported limitations or trade-offs. The paper positions AGSE-VNet as a VNet-based multimodal 3D framework using SE, AG, skip connections, and Categorical_Dice loss.

  • 2. Related Works: Traditional threshold-based segmentation is simple, but adaptive thresholding can fail when multiple organizational structures overlap.
  • 2. Related Works: Feature-rich machine-learning pipelines can cause dimensional disasters and waste substantial time when many features are input.
  • 2. Related Works: Deep-learning studies reported promising or strong results across BraTS and related medical-image segmentation tasks, including U-Net, residual, multiview, and group-convolution methods.
  • 2. Related Works: AGSE-VNet integrates SE and AG modules into VNet for end-to-end 3D MRI glioma segmentation using volumetric input and three-dimensional convolution.
  • 2. Related Works: SE weighting enhances useful feature channels, while AG filtering suppresses noise and irrelevant background and guides edge and structural information during decoding.
  • 2. Related Works: The model uses skip connections, Categorical_Dice loss, and BraTS 2020 evaluation, with the authors reporting good segmentation performance and clinical-trial potential.

3. Methodology

AGSE-VNet segments multi-sequence 3D brain-tumor MRI by combining SE modules in the encoder with AG modules in the decoder. The design selectively enhances useful channels, guides foreground and edge information, and uses encoder–decoder compression, decompression, and skip connections.

  • 3. Methodology: AGSE-VNet combines Squeeze-and-Excitation modules in encoder blocks with Attention Guided Filter modules in decoder blocks.The framework is proposed for segmenting multiple sequences of 3D brain-tumor MRI images.
  • 3.2. Squeeze-and-Excitation Blocks: SE modules recalibrate channel responses by modeling channel relationships to enhance useful information and suppress useless channels.The module uses global information and a squeeze–excitation process to adaptively adjust channel features.
  • 3.3 Attention Guided Filter Blocks: The Attention Guided Filter combines attention and guided filtering to generate high-resolution feature maps from low- and high-resolution inputs.It downsamples the guided feature map, estimates filter coefficients by minimizing reconstruction error, and upsamples them to produce the output.
  • 3.3 Attention Guided Filter Blocks: The attention mechanism is intended to highlight foreground features, reduce background influence, and improve segmentation accuracy.The framework also uses adjacent-layer feature reconstruction and cross-layer feature maps of different resolutions, and reports that performance is not affected by tumor size and location.
  • 3.4. Downsamplings: The network uses four encoder blocks for compression and decoder blocks for decompression, with skip connections supporting feature transfer between stages.Downsampling halves resolution and doubles channels, while decoder blocks receive features from corresponding encoder stages.
  • 3.3 Attention Guided Filter Blocks: AG modules replace VNet concatenation in the decoder while preserving input and output dimensions and enhancing features during reconstruction.Each decoder block includes upsampling, an AG module, and three convolutional layers.

4. Materials

The study uses multimodal BraTS2020 MRI data to segment nested brain-tumor regions, with preprocessing that standardizes modalities and divides cases into volumetric patches. Model performance is assessed using overlap, classification, sensitivity, and boundary-distance metrics.

  • Segmentation task: The task segments nested enhancement tumor, whole tumor, and tumor core regions from multimodal MRI data.Tumor masks include gangrene, edema, and enhancement areas, represented through nested labels.
  • Dataset: 369 training and 125 validation cases from BraTS2020 provide four MRI modalities and tumor masks for evaluation.The modalities are T1, T1-CE, T2, and FLAIR; validation masks are not used for training.
  • Training setup: Training uses an initial learning rate of 0.0001, dropout of 0.5, about 350,000 steps, and later learning-rate adjustment to 0.00003.The dataset is shuffled and repeatedly traversed to enhance robustness and generalization.

5. Results and Discussions

AGSE-VNet achieves strong brain-tumor segmentation performance, particularly for whole-tumor regions, across quantitative, visual, and noise-robustness evaluations. The results also identify weaker tumor-core and enhancing-tumor performance as an important limitation.

  • Quantitative results: 0.846, 0.849, 0.825, and 0.833 are reported for the training-set and validation-set Dice and Sensitivity measures, respectively.These values are described as significantly better than those of other methods.
  • Statistical analysis: Four-indicator scatter and box plots show higher Dice, Sensitivity, and Specificity levels, with few outliers and limited fluctuation.The reported indicators are Dice, Sensitivity, Specificity, and Hausdorff95 distance.
  • Qualitative results: Visual comparisons indicate excellent whole-tumor segmentation, while tumor-core predictions are slightly biased because the core features are small.The figures compare original images, labels, predicted subregions, and 3D segmentation results.
  • Discussion: The model segments tumors across different MRI intensities and tumor sizes or locations, extracting and segmenting tumor subregions with good performance.The authors state that attention recovers spatial information and fuses structural information from feature maps of different resolutions.
  • Quantitative results: AGSE-VNet performs well in whole-tumor segmentation, while comparison results show weaker performance in enhancing-tumor and tumor-core regions.The authors attribute the weaker ET performance partly to small, fuzzy targets that are difficult to extract.
  • Limitations and future work: The authors propose further region-of-interest segmentation and integration of additional algorithms to improve enhancing-tumor and tumor-core accuracy.These directions are presented as future work in response to the remaining regional accuracy differences.
  • Robustness analysis: AGSE-VNet produces similar segmentation results across noisy and noise-free inputs, indicating robustness when noise is present.The comparison covers the three tumor regions.

6. Conclusion

The conclusion presents AGSE-VNet as a 3D MRI brain-tumor segmentation method using SE-enhanced encoders, AG-filtered decoders, and a weighted Categorical_Dice loss. It reports BraTS2020 evaluation results while identifying ET and TC accuracy as remaining challenges.

  • 6. Conclusion: The method is reported to segment MRI images with different intensities and varying tumor sizes while extracting features from all three target regions.
  • 6. Conclusion: AGSE-VNet automatically segments the enhancing tumor, whole tumor, and tumor core in 3D MRI brain images.
  • 6. Conclusion: The architecture adds SE modules to encoder blocks and Attention Guide Filter blocks to decoder blocks without changing the network’s input-output size ratio.The SE module selects useful global information, while the decoder attention mechanism captures dependencies and enhances model performance.
  • 6. Conclusion: Categorical_Dice assigns background weight 0.1 and tumor-region weight 1 to address foreground-background voxel imbalance.
  • 6. Conclusion: BraTS2020 online verification evaluated the approach, which differed from top methods particularly on enhancing-tumor and tumor-core segmentation.The authors attribute the difficulty to the small, hard-to-extract features of these regions and identify improving their accuracy as future work.

Conflict of interest

The work reports support from Chinese national, provincial, medical, and visiting-scholar funding programs, alongside the AI for Health Imaging Award.

  • Conflict of interest: The study was funded in part by the National Natural Science Foundation of China and the Natural Science Foundation of Zhejiang Province.
  • Conflict of interest: Additional support came from Zhejiang’s Medical and Health Science and Technology Plan and its Domestic Visiting Scholar project.
  • Conflict of interest: The authors also received support from the AI for Health Imaging Award for the CHAIMELEON cancer-management project.

Authors' contributions

The authors divided responsibilities across study design, analysis, interpretation, writing, literature review, data collection, curation, and preparation of tables and figures.

  • Authors' contributions: XG, GY, and XL conceived and designed the study, analyzed and interpreted data, and contributed to writing the report.
  • Authors' contributions: XG, GY, JY, WY, XX, WJ, and XL contributed to the literature search, while JY and WY collected data.
  • Authors' contributions: XG, GY, XX, WJ, and XL performed data curation and contributed to the tables and figures; all authors approved the article.
Loading 2107.12046v1…