This is the official code base for Rhapso, a modular Python toolkit for the alignment and stitching of large-scale microscopy datasets.
Available on PyPI
- Summary
- Contact
- Supported Features
- Performance
- Layout
- Installation
- How To Start
- Try Rhapso on Sample Data
- Ray
- Run Locally w/ Ray
- Run on AWS Cluster w/ Ray
- Access Ray Dashboard
- Parameters
- Tuning Guide
- Build Package
Rhapso is still loading... and while we wrap up development, a couple things to know if you are outside the Allen Institute:
- This process requires a very specific XML structure to work.
- Fusion/Mutliscale is included but still under testing and development
Rhapso is a set of Python components used to register, align, and stitch large-scale, overlapping, tile-based, multiscale microscopy datasets. Its stateless components can run on a single machine or scale out across cloud-based clusters. Rhapso was developed by the Allen Institute for Neural Dynamics.
Questions or want to contribute? Please open an issue..
- Interest Point Detection - DOG based feature detection
- Interest Point Matching - Descriptor based RANSAC to match feature points
- Global Optimization - Align matched features between tile pairs globally
- Validation and Visualization Tools - Validate component specific results for the best output
- ZARR - Zarr data as input
- TIFF - TIFF data as input
- AWS - AWS S3 based input/output and Ray based EC2 instances
- Scale - Tested on 200 TB of data without downsampling
Rhapso/
└── Rhapso/
├── data_prep/ # Custom data loaders
├── detection/
├── evaluation/
├── fusion/
├── image_split/
├── matching/
├── pipelines/
│ └── ray/
│ ├── aws/
│ │ ├── config/ # Cluster templates (edit for your account)
│ │ └── alignment_pipeline.py # AWS Ray pipeline entry point
│ ├── local/
│ │ └── alignment_pipeline.py # Local Ray pipeline entry point
│ ├── param/ # Run parameter files (customize per run)
│ ├── interest_point_detection.py # Detection pipeline script
│ ├── interest_point_matching.py # Matching pipeline script
│ └── solver.py # Global solver script
├── solver/
└── visualization/ # Validation tools
# create and activate a virtual environment
python -m venv .venv && source .venv/bin/activate
# or: conda create -n rhapso python=3.10 && conda activate rhapso
# install Rhapso from PyPI
pip install Rhapso# clone the repo
git clone https://github.com/AllenNeuralDynamics/Rhapso.git
# create and activate a virtual environment
python -m venv .venv && source .venv/bin/activate
# or: conda create -n rhapso python=3.11 && conda activate rhapso
# install deps
pip install -r requirements.txtRhapso is driven by pipeline scripts.
- Each pipeline script has at minimum an associated param file (e.g. in
Rhapso/pipelines/ray/param/). - If you are running on a cluster, you’ll also have a Ray cluster config (e.g. in
Rhapso/pipelines/ray/aws/config/).
A good way to get started:
-
Pick a template pipeline script
For example:Rhapso/pipelines/ray/local/alignment_pipeline.py(local)Rhapso/pipelines/ray/aws/alignment_pipeline.py(AWS/Ray cluster)
-
Point it to your param file
Update thewith open("...param.yml")line so it reads your own parameter YAML. -
(Optional) Point it to your cluster config
If you’re using AWS/Ray, update the cluster config path. -
Edit the params to match your dataset
Paths, downsampling, thresholds, matching/solver settings, etc. -
Run the pipeline
The pipeline script will call the Rhapso components (detection, matching, solver, fusion) in the order defined in the script using the parameters you configured.
The quickest way to get familiar with Rhapso is to run it on a real dataset. We have a small (10GB) Z1 example hosted in a public S3 bucket, so you can access it without special permissions. It’s a good starting point to copy and adapt for your own alignment workflows.
XML (input)
- s3://aind-open-data/HCR_802704_2025-08-30_02-00-00_processed_2025-10-01_21-09-24/image_tile_alignment/single_channel_xmls/channel_488.xml
Image prefix (referenced by the XML)
- s3://aind-open-data/HCR_802704_2025-08-30_02-00-00_processed_2025-10-01_21-09-24/image_radial_correction/
Note: Occasionally we clean up our aind-open-data bucket. If you find this dataset does not exist, please create an issue and we will replace it.
This process has a lot of knobs and variations, and when used correctly, can work for a broad range of datasets.
First, figure out what type of alignment you need.
- Are there translations to shift to?
- If so, you’ll likely want to start with a rigid alignment.
Once you’ve run the rigid step, how does your data look?
- Did the required translations shrink to an acceptable level?
- If not, try again with new parameters, keeping the questions above in mind.
At this point, the translational part of your alignment should be in good shape. Now ask: are transformations needed? If so, you likely need an affine alignment next.
Your dataset should be correctly aligned at this point. If not, there are a number of reasons why, and we have listed some common recurrences and will keep this up to date.
There is a special case in some datasets where the z-stack is very large. In this case, you can use the split-dataset utility, which splits each tile into chunks. Then you can run split-affine alignment, allowing for more precise transformations without such imposing global rails.
Common Causes of Poor Alignment
- Not enough quality matches (adjust sigma threshold until you do)
- Data is not consistent looking (we take a global approach to params)
- Large translations needed (extend search radius)
- Translations that extend beyond overlapping span (increase overlap)
Interest Point Detection Performance Example (130TB Zarr dataset)
| Environment | Resources | Avg runtime |
|---|---|---|
| Local single machine | 10 CPU, 10 GB RAM | ~120 min |
| AWS Ray cluster | 560 CPU, 4.4 TB RAM | ~30 min |
*Actual times vary by pipeline components, dataset size, tiling, and parameter choices.*
Ray is a Python framework for parallel and distributed computing. It lets you run regular Python functions in parallel on a single machine or scale them out to a cluster (e.g., AWS) with minimal code changes. In Rhapso, we use Ray to process large scale datasets.
- Convert a function into a distributed task with
@ray.remote - Control scheduling with resource hints (CPUs, memory)
Tip
Ray schedules greedily by default and each task reserves 1 CPU, so if you fire many tasks, Ray will try to run as many as your machine advertises—often too much for a laptop. Throttle concurrency explicitly so you don’t overload your system. Use your machine's activity monitor to track this or the Ray dashboard to monitor this on your cluster:
- Cap by CPUs:
@ray.remote(num_cpus=3) # Ray will schedule each time 3 cpus are available - Cap by Memory and CPU if Tasks are RAM-Heavy (bytes):
@ray.remote(num_cpus=2, memory=4 * 1024**3) # 4 GiB and 2 CPU per task> - No Cap on Resources:
@ray.remote - Good Local Default:
@ray.remote(num_cpus=2)
Rhapso/Rhapso/pipelines/param/with open("Rhapso/pipelines/ray/param/your_param_file.yml", "r") as file:
config = yaml.safe_load(file)python Rhapso/pipelines/ray/local/alignment_pipeline.pyRhapso/pipelines/ray/param/with open("Rhapso/pipelines/ray/param/your_param_file.yml", "r") as file:
config = yaml.safe_load(file)Rhapso/pipelines/ray/aws/config/unified_yml = "your_cluster_config_file_name.yml"python Rhapso/pipelines/ray/aws/alignment_pipeline.pyTip
- The pipeline script is set to always spin the cluster down, it is a good practice to double check in AWS.
- If you experience a sticky cache on run params, you may have forgotten to spin your old cluster down.
This is a great place to tune your cluster's performance.
- Find public IP of head node.
- Replace the ip address and PEM file location to ssh into head node.
ssh -i /You/path/to/ssh/key.pem -L port:localhost:port ubuntu@public.ip.address - Go to dashboard.
http://localhost:8265
| Parameter | Feature / step | What it does | Typical range\* |
| :----------------- | :--------------------- | :-------------------------------------------------------------------------------------------- | :-------------------------------- |
| `dsxy` | Downsampling (XY) | Reduces XY resolution before detection; speeds up & denoises, but raises minimum feature size | 16 |
| `dsz` | Downsampling (Z) | Reduces Z resolution; often lower than XY due to anisotropy | 16 |
| `min_intensity` | Normalization | Lower bound for intensity normalization prior to DoG | 1 |
| `max_intensity` | Normalization | Upper bound for intensity normalization prior to DoG | 5 |
| `sigma` | DoG blur | Gaussian blur scale (sets feature size); higher = smoother, fewer peaks | 1.5 - 2.5 |
| `threshold` | Peak detection (DoG) | Peak threshold (initial min peak ≈ `threshold / 3`); higher = fewer, stronger peaks | 0.0008 - .05 |
| `median_filter` | Pre-filter (XY) | Median filter size to suppress speckle/isolated noise before DoG | 1-10 |
| `combine_distance` | Post-merge (DoG peaks) | Merge radius (voxels) to de-duplicate nearby detections | 0.5 |
| `chunks_per_bound` | Tiling/parallelism | Sub-partitions per tile/bound; higher improves parallelism but adds overhead | 12-18 |
| `max_spots` | Post-cap | Maximum detections per bound to prevent domination by dense regions | 8,0000 - 10,000 |
# Candidate Selection
| Parameter | Feature / step | What it does | Typical range |
| :----------------------------- | :------------------ | :---------------------------------------------------------------- | :------------- |
| `num_neighbors` | Candidate search | Number of nearest neighbors to consider per point | 3 |
| `redundancy` | Candidate search | Extra neighbors added for robustness beyond `num_neighbors` | 0 - 1 |
| `significance` | Ratio test | Strictness of descriptor ratio test; larger = stricter acceptance | 3 |
| `search_radius` | Spatial gating | Max spatial distance for candidate matches (in downsampled units) | 100 - 300 |
| `num_required_neighbors` | Candidate filtering | Minimum neighbors required to keep a candidate point | 3 |
# Ransac
| Parameter | Feature / step | What it does | Typical range |
| :---------------------------- | :------------------- | :---------------------------------------------------------------- | :------------- |
| `model_min_matches` | RANSAC | Minimum correspondences to estimate a rigid transform | 18 – 32 |
| `inlier_threshold` | RANSAC | Inlier tolerance scaling; larger = looser inlier threshold | 50 – 100 |
| `min_inlier_ratio` | RANSAC | Regularization strength during model fitting | 0.1 – 0.05 |
| `num_iterations` | RANSAC | Number of RANSAC trials; higher = more robust, slower | 10,0000 |
| `regularization_weight` | RANSAC | Weight applied to the regularization term | 0.05 - 1.0 |
| Parameter | Feature / step | What it does | Typical range |
| :----------------------- | :------------- | :----------------------------------------------------------------- | :------------------ |
| `relative_threshold` | Graph pruning | Reject edges with residuals above dataset-relative cutoff | 3.5 |
| `absolute_threshold` | Graph pruning | Reject edges above an absolute error bound (detection-space units) | 7.0 |
| `min_matches` | Graph pruning | Minimum matches required to retain an edge between tiles | 3 |
| `damp` | Optimization | Damping for iterative solver; higher can stabilize tough cases | 1.0 |
| `max_iterations` | Optimization | Upper bound on solver iterations | 10,0000 |
| `max_allowed_error` | Optimization | Overall error cap; `inf` disables hard stop by error | `inf` |
| `max_plateauwidth` | Early stopping | Stagnation window before stopping on no improvement | 200 |
| `regularization_weight` | RANSAC | Weight applied to the regularization term | 0.05 - 1.0 |
-
Start with Detection. The quality and density of interest points strongly determine alignment outcomes.
-
Target Counts (exaSPIM): ~25–35k points per tile in dense regions; ~10k for sparser tiles. Going much higher usually increases runtime without meaningful accuracy gains.
-
Inspect Early. After detection, run the visualization script and verify that peaks form clustered shapes/lines with a good spatial spread—a good sign for robust rigid matches.
-
Rigid → Affine Dependency. Weak rigid matches produce poor rigid transforms, which then degrade affine matching (points don’t land close enough). If tiles fail to align:
- Check match counts for the problem tile and its neighbors.
- Adjust high-impact detection knobs—
sigma,threshold, andmedian_filter—within sensible ranges. - Revisit
max_spotsandcombine_distanceto balance density vs. duplicate detections.
- Build the
.whlFile in the root of this repo:
cd /path/to/Rhapso
pip install setuptools wheel
python setup.py sdist bdist_wheelThe .whl file will appear in the dist directory. Do not rename it to ensure compatibility (e.g., rhapso-0.1-py3-none-any.whl).