AI4Life Calcium Imaging Denoising — Global Rank 2

Summary: A competitive research project focused on developing self-supervised machine learning models to denoise high-resolution biological calcium imaging — achieving Global Rank 2.

• Problem: High-resolution biological calcium imaging suffers from severe noise, and acquiring clean “ground truth” data for supervised machine learning is practically impossible in this domain.
• Solution: Built a self-supervised Noise2Void (N2V) model utilizing PyTorch for denoising high-resolution 3D calcium imaging data. The model extracts clean signals while strictly preserving the spatial and temporal integrity of the biological data. Evaluated efficacy using stSNR, PSNR, and SI-PSNR metrics.
• Tech Stack: Python, PyTorch, 3D Noise2Void, Docker, NumPy, SciPy.
• Outcome: Successfully containerized the final pipeline via Docker for deployment and evaluation on the Grand Challenge platform. Achieved Global Rank 2 by ensuring robust generalization across unseen datasets.

Denoising Pipeline Architecture

flowchart TD
    A["📥 Raw 3D TIFF Stack\n(Noisy Calcium Imaging)"] --> B["Preprocessing\n(Normalization + Patching)"]
    
    B --> C["3D Noise2Void\n(Self-Supervised)"]
    
    subgraph "N2V Training Loop (PyTorch)"
        C --> D["Blind-Spot Masking\n(Random Pixel Exclusion)"]
        D --> E["3D UNet Encoder-Decoder"]
        E --> F["Loss: MSE on\nMasked Pixels Only"]
        F -->|"Backprop"| E
    end
    
    E --> G["Denoised 3D Volume"]
    
    G --> H["Evaluation Metrics"]
    
    subgraph "Quality Assessment"
        H --> I["stSNR\n(Spatio-temporal SNR)"]
        H --> J["PSNR\n(Peak Signal-to-Noise)"]
        H --> K["SI-PSNR\n(Scale-Invariant)"]
    end
    
    G --> L["Docker Container\n(Reproducible Pipeline)"]
    L --> M["🏆 Grand Challenge\nSubmission Platform"]
    M --> N["Global Rank 2"]

• What I learned: Acquired advanced skills in handling massive 3D scientific datasets, tuning self-supervised machine learning models without ground truth labels, and strictly adhering to competitive research deployment standards using Docker.