Related Experiment Video
Updated: Jun 4, 2025
Co-culture of Glioblastoma Stem-like Cells on Patterned Neurons to Study Migration and Cellular Interactions
Published on: February 24, 2021
An interpretable multi-scale convolutional attention residual neural network for glioma grading with Raman
Qingbo Li1, Xupeng Shao1, Yan Zhou2
1School of Instrumentation and Optoelectronic Engineering, Precision Opto-Mechatronics Technology Key Laboratory of Education Ministry, Beihang University, Beijing 100191, China. qbleebuaa@buaa.edu.cn.
A new deep learning model, the Multi-Scale Convolutional Attention Residual Network (M-SCA ResNet), accurately classifies glioma grades using Raman spectroscopy. This advancement aids in personalized surgical planning and improves patient prognosis by distinguishing high-grade glioma, low-grade glioma, and normal tissue.
Area of Science:
- Biomedical Optics
- Machine Learning
- Oncology
Background:
- Glioma grading is crucial for personalized surgical planning and prognosis.
- Raman spectroscopy offers real-time glioma diagnosis but faces challenges in differentiating similar spectral characteristics between high-grade glioma (HGG), low-grade glioma (LGG), and normal tissues.
- Traditional machine learning methods struggle with accuracy due to spectral similarities, while deep learning requires careful network design for diverse spectral features.
Purpose of the Study:
- To propose a novel deep learning model, the Multi-Scale Convolutional Attention Residual Network (M-SCA ResNet), for accurate glioma grading.
- To enhance feature extraction capabilities for Raman spectral data by incorporating multi-scale attention and residual structures.
- To improve the classification accuracy of HGG, LGG, and normal tissues compared to conventional methods.
Main Methods:
- Development and application of the M-SCA ResNet, featuring multi-scale channel and spatial attention mechanisms with residual structures.
- Classification of HGG, LGG, and healthy tissue using the proposed M-SCA ResNet.
- Comparison of M-SCA ResNet performance against traditional machine learning and other neural network models.
- Utilized Gradient-weighted Class Activation Mapping (Grad-CAM) for visualization and interpretability of key Raman shifts.
Main Results:
- The M-SCA ResNet achieved an identification accuracy exceeding 85% for all three tissue types (HGG, LGG, normal).
- The model demonstrated the highest weighted F1-score among all tested methods.
- Grad-CAM analysis identified key Raman shifts correlating with biomolecular characteristics, validating the model's feature extraction capabilities and biological relevance.
Conclusions:
- The M-SCA ResNet effectively classifies glioma grades using Raman spectroscopy, outperforming conventional methods.
- The model's ability to extract relevant spectral features demonstrates a strong correlation with tissue biomolecular characteristics.
- This study validates the feasibility of M-SCA ResNet for in vivo and in situ glioma grading, supporting improved surgical and treatment planning.
