Related Experiment Video
Updated: Jun 29, 2025
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
Glioma segmentation based on dense contrastive learning and multimodal features recalibration
Xubin Hu1, Lihui Wang1, Li Wang1
1Engineering Research Center of Text Computing & Cognitive Intelligence, Ministry of Education, Key Laboratory of Intelligent Medical Image Analysis and Precise Diagnosis of Guizhou Province, State Key Laboratory of Public Big Data, College of Computer Science and Technology, Guizhou University, Guiyang 550025, People's Republic of China.
This study introduces DCL-MANet, a novel 3D deep learning model for accurate glioma segmentation from multimodal MRI. The model effectively disentangles multimodal features, improving segmentation of small and low-contrast lesions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuro-oncology
Background:
- Accurate glioma segmentation from multimodal Magnetic Resonance (MR) images is vital for diagnosis and grading.
- Existing methods struggle with segmenting small, low-contrast, or irregularly shaped lesions due to challenges in utilizing multimodal MR image information effectively.
Purpose of the Study:
- To propose a novel 3D glioma segmentation model, DCL-MANet, designed to overcome limitations in multimodal feature utilization for improved accuracy.
- To enhance the segmentation of challenging glioma regions, including small and low-contrast areas.
Main Methods:
- Developed a 3D glioma segmentation model, DCL-MANet, featuring multiple encoders for modality-specific feature extraction and a single decoder.
- Introduced a dense contrastive learning (DCL) strategy to disentangle multimodal semantic features into modality-specific and common components.
- Incorporated a feature recalibration block (RFB) with modality-wise attention to refine semantic features for better focus on relevant glioma characteristics.
Main Results:
- DCL-MANet demonstrated superior performance compared to state-of-the-art methods, with average improvements in Dice, Average Symmetric Surface Distance (ASSD), HD95, and Volumetric Similarity (Vs) across all tumor regions.
- Significant improvements were observed particularly in the segmentation of small enhancing tumor (ET) regions.
- Ablation studies confirmed the effectiveness of the DCL strategy and RFB, showing substantial gains in Dice and Vs, and reductions in ASSD and HD95 for the ET region when combined.
Conclusions:
- The DCL-MANet model effectively disentangles multimodal features and enhances modality-dependent semantics, offering a promising approach for accurate glioma segmentation.
- The proposed method shows significant potential for improving the segmentation of small and complex lesion regions in gliomas, aiding clinical diagnosis and treatment planning.
