Peak Detection of Mass Spectrometry Based on an Improved Light-Weight U-Net Model

Mass spectrometry is widely used in medical and other scientific fields because of its high analytical performance, and peak detection plays an important role in mass spectrometry data analysis. Many studies have proposed algorithms to address challenges such as noise interference and overlapping peaks, but most focus only on peak detection and neglect the calibration of complete characteristic peaks. This paper introduces a lightweight U-Net-based network that uses image segmentation techniques to perform peak detection on mass spectra. We use depthwise separable convolution to reduce model parameters and adjust the deep convolution kernel settings to enhance peak detection efficacy. Additionally, channel and spatial attention modules were integrated into the skip connections to improve network performance. The proposed model, trained on a custom dataset, demonstrated advantages in both model size and performance metrics compared to traditional U-Net models during testing. This approach provides a novel method for peak detection in mass spectrometry.