MAPP-Net: Efficient Mamba-based Particle Picking for Cryo-ET
Zhang, Zuotong
Zhang, Zuotong
Author
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Accurate and efficient particle picking remains a crucial challenge in the analysis of cryo-electron tomography (cryo-ET) data due to low signal-to-noise ratios, densely packed environments, and high computational costs. In this work, we propose MAPP-NET, a lightweight and high-performance deep learning framework for 3D particle detection in cryo-ET tomograms. MAPP-NET leverages a stack of Vision State Space Blocks (VSSBs) to model long-range dependencies in a computationally efficient manner, replacing traditional attention-based components with a state-space formulation. To enhance sensitivity to true particles, we introduce a modified classification loss that penalizes false negatives more heavily. Additionally, we aggregate multi-scale features from each VSSB block to enrich the final representation passed to the prediction heads. Comprehensive experiments on benchmark datasets—including SHREC2020, SHREC2021, EMPIAR-10045, and EMPIAR-10651—demonstrate that MAPP-NET achieves competitive or superior performance compared to state-of-the-art baselines, while maintaining a significantly lower number of parameters and faster inference times. Ablation studies confirm the effectiveness of multi-directional flattening and VSSB blocks. These results establish MAPP-NET as a practical and scalable solution for highthroughput cryo-ET analysis. Future improvements will explore the integration of positional embeddings and adaptive feature map configurations to further enhance generalization across tomographic conditions.
Citation
Zuotong Zhang, “MAPP-Net: Efficient Mamba-based Particle Picking for Cryo-ET,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Cryo-ET, Object Detection/Particle Picking, State Space Model, Light Weight Model