An Interpretable Multi-Scale Convolutional Attention Residual Neural Network for Glioma Grading with Raman Spectroscopy

Since the malignancy of gliomas varies with their grade, classifying gliomas of different grades can assist doctors in developing personalized surgical plans during surgery, thereby improving the prognosis. Raman spectroscopy is an optical method for real-time glioma diagnosis. However, high-grade glioma (HGG, WHO grades III and IV), low-grade glioma (LGG, WHO grades I and II) and normal tissue have similar biochemical components, leading to similar spectral characteristics. This similarity reduces classification accuracy when using traditional machine learning methods. In contrast, deep learning offers enhanced feature extraction capabilities without the need for extensive feature engineering. Nevertheless, the diversity in the scale of spectral features presents challenges in designing a neural network that effectively adapts to these characteristics. To address these issues, this paper proposes a Multi-Scale Convolutional Attention Residual Network (M-SCA ResNet), which incorporates multi-scale channel and spatial attention mechanisms along with residual structures to improve the model's feature extraction capabilities. The algorithm presented in this study, was employed to classify HGG, LGG, and healthy tissue and was compared with conventional machine learning and neural networks. The results indicate that the M-SCA ResNet achieved an identification accuracy exceeding 85% for all three tissue types, along with the highest weighted