Automated red blood cell segmentation in low-grade glioma Raman imaging using a hybrid wavelet-transformer network

This study aims to develop an efficient and accurate deep learning method for automated red blood cell (RBC) segmentation from Low-Grade Glioma (LGG) Raman spectroscopic imaging data. Accurate segmentation is crucial for enhancing the reliability of Raman analysis, as RBC signals can interfere with the precise identification of tumor characteristics. To this end, we propose a novel hybrid neural network architecture named FusedWaveletUNet. This model, built upon the U-Net framework, innovatively integrates Discrete Wavelet Transform (DWT/IWT) to capture multi-scale frequency domain information, introduces a Transformer module for global contextual modeling, and leverages a Bottleneck Attention Module (BAM) to optimize feature propagation. Experimental results demonstrate that our model outperforms various mainstream segmentation networks across key metrics on a dataset comprising 500 image samples, achieving an IoU of 0.81 and a Dice coefficient of 0.89. The proposed FusedWaveletUNet offers a robust automated solution for RBC segmentation in LGG Raman imaging, promising to significantly improve the analytical quality of medical Raman imaging and provide more precise information for clinical diagnosis and biological research.