Context-Aware Network Study Report for Glioma Segmentation

Glioma is a common type of adult brain tumor that severely harms health and has a high mortality rate. To provide sufficient evidence for early diagnosis, surgical planning, and postoperative observation, multimodal Magnetic Resonance Imaging (MRI) has been widely applied in this field. The objective of this study is to incorporate contextual information in the automated segmentation of brain glioma, which provides fundamental clues when dealing with local ambiguity.

Research Background

Previous studies have shown that deep neural network-based methods in brain glioma segmentation are promising. However, these methods lack a strong strategy to combine the contextual information of the tumor cells and their surroundings. Although existing automatic segmentation methods have improved accuracy, they still suffer from the issue of local ambiguity and have not fully considered the relationship between tumor cells and their environment.

Paper Source

This study was jointly authored by Zhihua Liu, Lei Tong, Long Chen, Feixiang Zhou, Zheheng Jiang, Qianni Zhang, Yinhai Wang, Caifeng Shan, Ling Li, and Huiyu Zhou, and was published in the journal “IEEE Transactions on Medical Imaging” in July 2021.

Research Process

The paper proposes a novel approach called the Context-Aware Network (Canet) for glioma segmentation. The main workflow includes the following steps:

1. Data Preparation

The study used publicly accessible brain glioma segmentation datasets BRATS2017, BRATS2018, and BRATS2019. These datasets contain multimodal MRI images, including sequences of T1, T1 post-contrast enhancement (T1CE), T2, and Fluid-Attenuated Inversion Recovery (FLAIR).

2. Designing the Overall Network Architecture

• Encoder and Decoder: A structure with an encoder and decoder featuring skip connections is used. The encoder reduces the spatial dimensions of the feature maps, while the expansion path restores the spatial dimensions of the feature maps and recovers the details of the targets.
• Feature Interaction Map: A feature interaction map is constructed to capture the long-distance relationships between feature nodes, utilizing graph convolution to transmit and update node features.

3. Feature Fusion

A feature fusion model based on conditional random fields, called Context-Guided Attentive Conditional Random Field (CGA-CRF), is designed to effectively learn optimal latent features for the final segmentation. CGA-CRF uses the mean-field approximation method and formulates it as a convolution operation, allowing it to be seamlessly integrated with any neural network for end-to-end training.

4. Context-Guided Feature Extraction

• Graph Context: Projection and adaptive sampling to learn the relationships between feature nodes by building an interaction graph.
• Convolutional Context: Use of the encoder-decoder structure with deep supervision mechanisms to improve training.

5. Experimental Evaluation

Extensive experimental evaluations were conducted, comparing the proposed method with existing state-of-the-art methods across different segmentation metrics. Experimental results showed that the proposed algorithm has superior or competitive segmentation performance on the training and validation sets compared to several state-of-the-art methods.

Main Results

The evaluation on the BRATS datasets yielded the following principal results:

• Dice Score: Canet achieved Dice scores of 0.903 and 0.873 for whole tumor and tumor core, respectively, significantly higher than other methods.
• Hausdorff95 Distance: Canet’s Hausdorff95 distances for whole tumor and tumor core were 3.569 and 4.036, respectively, both lower than those of other methods.

Conclusion

The new brain glioma segmentation method proposed in this paper significantly improves segmentation accuracy and performance by introducing a feature interaction map combined with contextual information. The main contributions include:

1. Proposing a novel brain glioma segmentation approach that treats the feature interaction map as a parallel auxiliary branch to model the relationship between glioma cells and their environment.
2. The customized CGA-CRF framework utilizes and aggregates intermediate feature representations further.
3. Extensive evaluations indicate that our proposed method outperforms several state-of-the-art methods on multimodal brain tumor image segmentation challenge datasets (BRATS2017, BRATS2018, and BRATS2019).

The experimental results show that the proposed Context-Aware Network (Canet) performs excellently in segmentation tasks. Future research could consider integrating new training methods to better address the imbalance issues within the datasets.