hisalign 0.2.0

HE-to-IHC whole-slide image alignment pipeline

Whole-slide image alignment for H&E and multiplex IHC markers

English · 简体中文

"Align once. Map everywhere."
"一次配准，处处映射"

---

✨ Why HISAlign

Problem	Solution
H&E and multiple IHC slides of the same tissue are spatially misaligned	Rigid + optical-flow non-rigid registration, per-marker alignment to H&E
Registration results depend on open slide handles and are hard to reuse	Only numpy arrays and transform parameters are saved; fully serializable
Downstream analysis needs to map ROIs from H&E to IHC	Offline warp_xy interface: provide level-0 coordinates and get mapped coordinates
Need a quick way to assess registration quality	Optional unified viz_report.html with slide summary and patch gallery

---

🎨 Workflow

flowchart LR
    HE["H&E Reference"]
    IHC["IHC Markers\nCD3 / Ki67 / ..."]
    PRE["Optical Density\nPreprocessing"]
    RIGID["Rigid Registration"]
    NR["Optical-Flow\nNon-Rigid"]
    MODEL["model.pkl\nSerializable Model"]
    WARP["warp_xy\nOffline Mapping"]
    VIZ["viz_report.html"]

    HE --> PRE
    IHC --> PRE
    PRE --> RIGID
    RIGID --> NR
    NR --> MODEL
    MODEL --> WARP
    NR --> VIZ

---

🚀 Installation

Install from PyPI using uv:

uv tool install hisalign

Or install it as a library in your current environment:

uv pip install hisalign

Verify the CLI entry point:

hisalign --help

---

🧪 Quickstart

Python API

from hisalign import HisAlign, HisAlignModel

# 1. Fit and save the model
aligner = HisAlign(
    he_path="HE.kfb",
    ihc_paths={"CD3": "CD3.svs", "Ki67": "Ki67.svs"},
    registration_level=3,
    max_image_dim_px=1024,
    preprocessing="od",
    feature_detector="kaze",
    mpp=0.25,  # explicit when slide metadata lacks MPP
)
model = aligner.fit()
model.save("model.pkl")

# 2. Offline coordinate mapping (no slides needed)
loaded = HisAlignModel.load("model.pkl")
mapped = loaded.warp_xy(
    coords=[[1000, 2000]],  # H&E level-0 pixel coordinates
    marker="CD3",
    direction="he_to_ihc",
)
print(mapped)  # -> [[ihc_x, ihc_y]]

CLI

1. Register and save the model

hisalign register \
  --he HE.kfb \
  --ihc CD3=CD3.svs \
  --ihc Ki67=Ki67.svs \
  --output model.pkl \
  --config configs/default.yaml \
  --mpp 0.25

If you omit marker=, the marker name is derived from the last token of the filename stem. hisalign register only produces model.pkl by default; set generate_report: true in the config to also emit viz_report.html.

2. Warp coordinates with the model

hisalign warp \
  --model model.pkl \
  --marker CD3 \
  --direction he_to_ihc \
  --coords coords.csv \
  --output mapped.csv

The input CSV must contain x and y columns; the output adds marker and direction columns.

3. Generate visualizations from an existing model

hisalign visualize \
  --model model.pkl \
  --output-dir ./out \
  --config configs/default.yaml

Default behavior: hisalign register only writes model.pkl by default. Set generate_report: true in your config (or in configs/default.yaml) to also produce the unified viz_report.html.

register also writes viz_report.html automatically when visualization is enabled.

---

🖼️ Works with Plain Images Too

The repo includes whole-slide thumbnail exports of an H&E slide and a CD3 IHC slide (examples/images/he.jpg and examples/images/ihc.jpg) for a quick demo:

python examples/register_jpg.py --output-dir ./out

For synthetic data:

python examples/register_jpg.py --synthetic --output-dir ./out

Or provide your own images:

python examples/register_jpg.py --he he.jpg --ihc ihc.jpg --output-dir ./out

---

📐 Coordinate Convention

⚠️ All coordinates are level-0 (highest resolution) pixel coordinates.

• Order is (x, y) = (column, row).
• Origin is the top-left corner.
• For coordinates from another pyramid level, scale them to level-0 first.

---

📦 Supported Formats

Type	Formats
KFBio native	.kfb
OpenSlide	.svs, .tif/.tiff, .ndpi, .vms/.vmu, .mrxs, .scn
Plain images	.jpg, .jpeg, .png, .bmp

---

📤 Outputs

After running hisalign register:

• model.pkl — serializable alignment model containing all spatial transforms; no original slides required afterwards.

If visualization is enabled, the same directory also contains:

• viz_report.html — unified visualization report: a single self-contained web page with two tabs. The "Slide Summary" tab shows overlay comparisons, rTRE statistics, per-marker thumbnails, and deformation fields. The "Patch Gallery" tab shows the sampled patches with global context and local HE/IHC patch comparisons. Images are JPEG-compressed and web-sized so the file stays small enough to open and share.

Tip: The default configs/default.yaml has generate_report: false. Set it to true to get viz_report.html directly from hisalign register without running hisalign visualize.

---

⚙️ Configuration

Key items in configs/default.yaml:

registration_level: 3          # pyramid level used for registration
max_image_dim_px: 1024         # longest side of the processed registration image
preprocessing: "od"            # "od" optical density | "gray" simple grayscale
feature_detector: "kaze"       # kaze / akaze / sift / orb / brisk
feature_n_levels: 3
match_max_ratio: 1.0           # Lowe ratio test; 1.0 disables
mpp: null                      # level-0 pixel size (µm/px); set when metadata is missing

# Visualization
viz_sample_n: 8              # number of patches in gallery, 0 to disable
viz_global_thumb_max_dim_px: 512  # longest side of global location thumbnails
viz_image_format: "jpeg"      # png | jpeg — base format for most figures
viz_image_quality: 85         # JPEG quality (0-100), ignored for PNG
viz_overlay_dpi: 80           # DPI for overlay figures
viz_thumb_dpi: 80             # DPI for per-marker thumbnails
viz_deformation_dpi: 60       # DPI for deformation field plots
viz_patch_dpi: 80             # DPI for patch gallery figures
viz_patch_image_format: "jpeg"  # format specifically for patch gallery
viz_patch_image_quality: 85     # JPEG quality for patch gallery
viz_patch_max_px: 512         # local patches are resized to this longest side before plotting
viz_patch_col_inches: 2.5     # width per marker column in patch gallery figures
generate_report: false        # whether to generate viz_report.html
report_rtre_threshold: 5.0    # rTRE threshold for "good" highlighting

---

🗂️ Project Structure

hisalign/
├── README.md
├── README.zh.md
├── pyproject.toml
├── configs/default.yaml
├── examples/
│   ├── register_jpg.py
│   └── images/
│       ├── he.jpg
│       └── ihc.jpg
├── src/hisalign/
│   ├── api.py
│   ├── cli.py
│   ├── preprocessing.py
│   ├── registration/
│   ├── slide_io/
│   └── viz.py
└── tests/

---

🧪 Try It

The fastest way to verify the installation and see HISAlign in action is to run the bundled example:

python examples/register_jpg.py --output-dir ./out

This registers the real whole-slide thumbnails in examples/images/ and produces:

• out/model.pkl
• out/00_unregistered.png
• out/01_rigid.png
• out/02_nonrigid.png

Example Results

After running the command above, open the generated overlays:

• out/00_unregistered.png — green/magenta overlay before registration (structures are shifted).
• out/01_rigid.png — overlay after rigid registration.
• out/02_nonrigid.png — overlay after non-rigid registration; overlapping structures turn white/gray.

Green = H&E, magenta = CD3 IHC.

---

🙋 Author

Created with 💙 by Yifan Feng
📧 [e-mail]

---

📚 References

• VALIS: Virtual Alignment of pathology Image Series
• DISK: Learning local features with policy gradient (Tyszkiewicz et al., NeurIPS 2020)
• LightGlue: Local Feature Matching at Light Speed (Lindenberger et al., CVPR 2023)

---

Making whole-slide alignment as intuitive as solving a jigsaw puzzle.