Janani Vaiyapuriappan, Phoebe Wu, Siyang Qin
Department of Biomedical Engineering, Whiting School of Engineering, Johns Hopkins University
Supervised by Dr. Adam Charles · Acknowledged: Gabrielle Coste
Presented at JHU BME Design Day 2026

Status: Core pipeline (restoration + segmentation + GUI) complete and validated. Presented at JHU BME Design Day, April 2026. Longitudinal synapse tracking and extended cross-dataset validation are in active development.

Synpipe is a modular, containerized computational pipeline that integrates deep learning–based image restoration and synapse segmentation into a single reproducible framework, with a live GUI dashboard for real-time monitoring on HPC infrastructure.

Modern in vivo synaptic imaging generates large, noisy 3D two-photon microscopy datasets that require multiple sequential processing steps — image restoration, segmentation, and longitudinal tracking. Existing tools like XTC (cross-modality image restoration) and MR-CNN (synapse segmentation) operate in entirely separate software environments with incompatible Python, CUDA, TensorFlow, and PyTorch dependencies, making unified, reproducible analysis difficult.

Synpipe solves this by containerizing each tool independently using Docker and orchestrating the full pipeline with Hydra, eliminating dependency conflicts and enabling one-command execution across heterogeneous computing environments — local machines, HPC clusters, and cloud.

• SSIM ≥ 0.99 and Pearson r ≈ 1.00 across all output types (XTC-processed images, MR-CNN segmentations, overlays), confirming that containerization preserves computational fidelity
• Zero spatial shift detected across all comparisons
• Hydra pipeline SSIM = 0.9999 on XTC output — statistically identical to original workflow
• Live GUI dashboard running on navon A100 GPU cluster, accessible via browser at localhost:5000
• Real-time log streaming — researchers can monitor XTC patch processing progress without terminal access

Raw 3D 2P In Vivo Volume
        │
        ▼
┌──────────────────┐
│  GUI Initiation  │  ← User loads volumetric data, triggers workflow, manages parameters
└────────┬─────────┘
         │
         ▼
┌──────────────────────────────────────────────────┐
│           XTC Restoration Module                 │
│  • Reimplemented 3D U-Net XTC                    │
│  • Optimized sparse-array processing             │
│  Output: Restored Volume (High SNR, Super-Res)   │
└────────┬─────────────────────────────────────────┘
         │
         ▼
┌──────────────────────────────────────────────────┐
│           MR-CNN Segmentation Module             │
│  • Adapted ResNet-101 FPN with Group Norm        │
│  • Volumetric + slice-by-slice normalization     │
│  Output: Segmented Synapses (3D Instance Masks)  │
└────────┬─────────────────────────────────────────┘
         │
         ▼
┌──────────────────────────────────────────────────┐
│           Tracking System                        │
│  • LAP-based synapse linking across timepoints   │
│  • Biologically constrained (3.0 µm/frame)       │
│  • Outputs tracks.csv with trajectory data       │
└────────┬─────────────────────────────────────────┘
         │
         ▼
┌──────────────────┐
│  GUI & Validation│  ← Interactive QC/Manual Correction, statistical report, data export
└──────────────────┘

Each module runs in its own isolated Docker container. Hydra orchestrates execution order, parameter propagation, and data flow between stages via mounted volumes.

Both environments were fully upgraded and containerized. Due to incompatible TensorFlow/PyTorch dependency stacks, XTC and MR-CNN are maintained as separate containers — a core design decision that enables modular, conflict-free deployment.

Containerized and Hydra-orchestrated pipeline outputs were compared against the original non-containerized workflow using three quantitative metrics:

Containerized Run Comparison

Hydra Pipeline Comparison

All outputs maintained identical 1024×1024 dimensions with no resizing. Minor numerical differences are attributable to expected floating-point precision variability, not algorithmic discrepancy.

The full pipeline is defined in a single YAML config file, enabling parameter changes without modifying code:

pipeline:
  input_dir: /path/to/input
  output_root: /path/to/results
  run_timestamp: 2026-04-11_09-26-50

stages:
  xtc:
    enabled: true
    image: xtc:synpipe
    xy_resolution: 0.095
    z_resolution: 1.0

  mrcnn:
    enabled: true
    image: mrcnn:synpipe

  overlay:
    enabled: true

docker:
  use_gpus: true
  gpus: all
  run_as_current_user: true

Variable interpolation (e.g., ${pipeline.output_root}) ensures outputs are organized consistently across runs without manual path editing.

The full Synpipe pipeline is available in this repository, including Docker container configurations, Hydra orchestration scripts, Flask GUI dashboard, and LAP tracking module.

See synapsepipe/README.md for GUI setup instructions.

See Hydra/README.md for command-line pipeline usage.

Synpipe was built to address a real bottleneck in neuroscience research. Synaptic connections are the fundamental basis of neural communication — their dynamic localization and density at synapses regulate synaptic strength and plasticity, mediated by AMPA receptors (AMPARs). Recent transgenic models (SEP–GluA2) enable direct in vivo visualization of these dynamics using two-photon microscopy, generating large longitudinal 3D datasets.

The primary contribution of Synpipe is not new algorithms but reproducible computational infrastructure: transforming a fragmented, error-prone workflow into a unified, portable, one-command system that any research group can deploy.

The current implementation covers image restoration, segmentation, and the GUI dashboard. The following are pending or planned:

Planned

• Docker Compose for single-command full-stack deployment
• Extended validation across diverse imaging datasets and species
• Export in standard neuroscience formats (SWC, NIfTI)
• Cross-dataset benchmarking of runtime performance at scale

• Chen Z, et al. Automatic detection of fluorescently labeled synapses in volumetric in vivo imaging data. bioRxiv (2025). doi: 10.1101/2025.01.22.634278
• Xu YKT, Graves AR, Coste GI, et al. Cross-modality supervised image restoration enables nanoscale tracking of synaptic plasticity in living mice. Nature Methods 20, 935–944 (2023).
• Graves AR, et al. Visualizing synaptic plasticity in vivo by large-scale imaging of endogenous AMPA receptors. eLife 10:e66809 (2021).
• Docker Inc. Docker Documentation. https://docs.docker.com/
• Meta AI. Hydra Documentation. https://hydra.cc/

MIT License