Multi-Level Feature Exploration and Fusion Network for Prediction of IDH Status in MRI

Background

Glioma is the most common malignant primary brain tumor in adults. According to the 2021 World Health Organization (WHO) classification of tumors, genotype plays a significant role in the classification of tumor subtypes, especially the isocitrate dehydrogenase (IDH) genotype, which is critically important in the diagnosis of gliomas. Clinical studies have shown that gliomas with IDH mutations are driven by specific epigenetic variation characteristics that affect enzyme activity, cellular metabolism, and biological properties. Compared to gliomas with wild-type IDH, those with IDH mutations are more sensitive to temozolomide and have a better prognosis. Currently, the determination of IDH status mainly relies on genetic sequencing or immunohistochemical analysis of tissue specimens after invasive surgery. However, invasive procedures may delay the final treatment decision and even lead to tumor metastasis. Therefore, there is an urgent need for non-invasive methods to predict IDH status preoperatively, to help formulate appropriate treatment plans for glioma patients.

Paper Source

This paper was published in the IEEE Journal of Biomedical and Health Informatics, titled “Multi-Level Feature Exploration and Fusion Network for Prediction of IDH Status in Gliomas from MRI”, and was published in January 2024. The authors of the paper include Jiawei Zhang, Jianyun Cao, Fan Tang, Tao Xie, Qianjin Feng, and Meiyan Huang, who are from the School of Biomedical Engineering, Southern Medical University, Zhujiang Hospital, and Southern Hospital.

Research Process

Design of Multi-Level Feature Exploration and Fusion Network (MFEFNet)

This study proposes a new method named Multi-Level Feature Exploration and Fusion Network (MFEFNet) to explore features associated with glioma IDH status and accurately predict them by combining multiple features. The specific research process is as follows:

1. Segmentation-guided feature extraction module (SFE): This module guides the network to extract features highly relevant to the tumor by incorporating segmentation tasks. Using ResNet50 as the encoder and adding channel adaptive weights, the network extracts features at different levels.

2. Asymmetry magnification module (AMF): This module detects T2-FLAIR mismatch signs and their related features, significantly enhancing feature representation through both image-level and feature-level difference amplification. The process includes element-wise subtraction at the image level to highlight the differences between T2 and FLAIR sequences and extracting mismatch features using a Siamese structure.

3. Dual-attention feature fusion module (DFF): This module includes self-attention mechanisms and Multi-instance Learning (MIL) attention pooling to combine and utilize relationships between different features. Features are extracted from each 2D slice of the patient and then fused considering both intra-slice and inter-slice feature information for final IDH prediction.

Research Experiments

The study was evaluated on a multi-center dataset, showing promising performance on an independent clinical dataset. Specific experimental methods include:

1. Data Preprocessing: All MRI images underwent bias field correction, registration to T1 MRI images, skull stripping, and interpolation to a voxel resolution of 0.75×0.75 mm. Intensity normalization was performed, and images were cropped to a size of 224×224 for network input.
2. Network Training: Using the PyTorch framework and an NVIDIA 12GB Pascal Titan X GPU, five-fold cross-validation was employed on the training set, with a maximum of 200 epochs, the Adam optimizer, and a one-cycle learning rate.
3. Performance Evaluation: Metrics such as the area under the ROC curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE) were used to evaluate the IDH prediction performance.

Research Results

Research Results demonstrated the effectiveness of the three modules:

1. SFE module: Incorporating the segmentation task improved IDH prediction performance, indicating that tumor-related features could be better extracted.
2. AMF module: Differential features enhanced the precision of IDH prediction, particularly when combining T2-FLAIR mismatch features, significantly boosting feature representation capability.
3. DFF module: Compared to simple feature concatenation, the multi-head self-attention mechanism and multi-instance learning better captured inter-slice feature correlations, improving prediction performance. Additionally, comparative validation showed that the MFEFNet method outperformed other methods in prediction and segmentation performance, more effectively extracting potential information related to the tumor.

Research Value

Scientific Value

The proposed method provides a new non-invasive approach for preoperative IDH prediction by combining segmentation guidance, multi-level difference magnification, and dual-attention mechanisms, improving the performance of IDH status prediction.

Application Value

The performance of this method on actual clinical datasets demonstrates its good generalization ability, showing promise for clinical practice to provide important references for the diagnosis and treatment of glioma patients.

Research Highlights

• Innovativeness: By designing a multi-level feature exploration and fusion network that leverages T2-FLAIR mismatch features and self-attention mechanisms, it effectively captures features related to IDH status.
• Effectiveness: Extensive validation on multi-center data demonstrates the superiority and good predictive performance of the algorithm.

The new deep learning method proposed in this study provides an effective solution for non-invasive IDH status prediction, showing significant application prospects in glioma diagnosis and treatment planning.