Work

Enhancing Mass Spectrometry Imaging : A Deep Learning Based Approach

Year: 2024
Type: dissertation
Abstract: Mass spectrometry (MS) is essential for determining molecular compositions by measuring mass-to-charge ratios. When combined with imaging, Mass Spectrometry Imaging (MSI) spatially maps these distributions, useful in fields like biomedical research. However, MSI often has lower spatial resolution and longer acquisition times than traditional microscopy. This study explores deep learning models to enhance MSI resolution and reduce acquisition times while maintaining image quality. We evaluated Fast Super-Resolution Convolutional Neural Network (FSRCNN), Efficient Sub-Pixel Convolutional Neural Network (ESPCN), and Real-Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN). These models were trained on MSI and optical images from the Lanekoff Lab and public datasets. Metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) assessed image quality. Results showed that CNN-based models performed consistently, while Real-ESRGAN achieved higher PSNR but lower SSIM values due to artifacts. We also explored fusing MSI data with high-resolution optical images using U-net architectures and Spatial Transformer Networks (STNs) for alignment and fusion. The STN method outperformed traditional MATLAB-based alignment methods, achieving higher mutual information (MI), normalized mutual information (NMI), and lower mean square error (MSE). The U-net-based fusion effectively combined structural details from optical images with the chemical specificity of MSI data. Overall, deep learning techniques significantly enhance MSI image resolution and quality. Future work should optimize these models for better efficiency and explore hybrid approaches to further enhance MSI and optical image integration.
Language: English
Open Access status: green
OpenAlex ID: W7116819357
Cites:
Cited by:
Related to: 9