To address the challenges in mass spectrometry data processing, including poor detection of low-abundance metabolites, severe noise interference, and the reliance on labor-intensive parameter tuning in conventional peak detection methods, a deep learning-based peak detection algorithm termed MSPeakNet was developed. The model used a multi-scale convolutional module and a U-net encoder-decoder architecture integrated with an attention mechanism, enabling high-precision peak identification and noise reduction in complex biological matrices. MSPeakNet achieved precision and F1 scores comparable to CWT, with superior noise robustness, while maintaining competitive recall. The developed method was subsequently applied to the in vivo metabolite profiling of total triterpenoids from Eleutherococcus senticosus leaves (ESLT) in an MPTP-induced mouse model of Parkinson's disease (PD) using UPLC-Q-TOF/MS analysis. A total of 143 ESLT-related compounds, comprising 88 prototypes and 55 metabolites, were identified in plasma, urine, feces, and brain tissue, of which 16 prototypes were detected in brain tissue. Pharmacodynamic evaluation further confirmed that ESLT significantly ameliorated MPTP-induced motor dysfunction and dopaminergic neuronal loss. Comparative analysis further revealed distinct profiles of ESLT-related differential compounds between PD model and normal mice across multiple biological matrices; 14, 9, 6, and 4 differential metabolites were identified in feces, urine, plasma, and brain, respectively, suggesting that disease status may affect the biotransformation, distribution, and elimination of ESLT constituents. A robust and efficient deep learning-based approach for mass spectrometry peak detection was thereby established, and critical experimental evidence was provided for the discovery of anti-Parkinsonian active ingredients and quality markers from E. senticosus leaves.