TIAToolbox v2.0.0 (2026-03-11)

✨ Major Updates and Feature Improvements

⚙️ Engine Redesign (PR #578)

TIAToolbox 2.0.0 introduces a completely re-engineered inference engine designed for significant performance, scalability, and memory-efficiency improvements.

Key Enhancements

• A modern processing stack built on Dask (parallel/distributed execution) and Zarr (chunked, out-of-core storage)
• Standardised output formats across all engines:
  • Python dict
  • Zarr
  • AnnotationStore (SQLite-backed)
  • QuPath JSON
• Cleaner runtime behavior with reduced warning noise and a unified progress bar
• More predictable memory usage through chunked streaming
• Broader test coverage across engine components

🗺️ Improved QuPath Support

Enhancements include:

• Better handling of GeoJSON
• Support for multipoint geometries (#841)
• Improved semantic output helpers:
  • dict_to_store_semantic_segmentor (#926)
  • OME-TIFF probability overlays (#929)

🔬 New Nucleus Detection Engine

A dedicated nucleus detection pipeline has been added, built on the redesigned engine for improved accuracy and efficient large-scale processing.

🧠 KongNet Model Family

TIAToolbox 2.0.0 introduces KongNet, a high-performance architecture that achieved top results across multiple international challenges:

• 🥇 1st place: MONKEY Challenge (overall detection)
• 🥇 1st place: MIDOG (mitosis detection)
• ⭐ Top-tier performance on PUMA

Multiple pretrained variants are available (CoNIC, PanNuke, MONKEY, PUMA, MIDOG), each with standardised IO configurations.

🧬 Expanded Foundation Model Support

Additional foundation models are now supported (#906), broadening the range of high-capacity architectures available for feature extraction and downstream tasks.

🖼️ SAM Segmentation in TIAViz

TIAViz now integrates Meta’s Segment Anything Model (SAM), enabling:

• Interactive segmentation
• Rapid region extraction
• Exploratory annotation workflows

Simplified SAM usage (#968) streamlines its integration into analysis pipelines.

🖼️ WSI Registration Visualization in TIAViz

TIAViz now supports interactive WSI registration visualisation, allowing users to compare aligned slides in two modes:

• Side‑by‑side view — view fixed and moving WSIs next to each other for direct comparison
• Overlay mode — blend registered WSIs with adjustable transparency for visual inspection of alignment quality

This feature enables intuitive, high‑resolution exploration of slide registration results directly within TIAViz.

🧩 Enhanced WSIReader & Metadata Handling

Major improvements include:

• More robust cross-vendor metadata extraction (#1001)
• Multichannel image support (PR #825) for immunofluorescence and non-RGB modalities
• Simplified Windows installation using openslide-bin (no manual DLL steps)
• macOS Tileserver fix (#976)
• Improved DICOM reading (#934)

☁️ New Cloud-Native Reader: FsspecJSONWSIReader (PR #897)

A new reader supporting fsspec-compatible filesystems, enabling seamless access to WSIs stored on:

• S3
• GCS
• Azure
• HPC clusters
• Any fsspec-supported backend

This enables cloud-native and distributed data workflows.
Contributed by @aacic

🤗 Pretrained Models Migrated to Hugging Face

All pretrained models and sample assets have been migrated (#945, #983), improving:

• Download reliability
• Versioning and reproducibility
• Caching and CI integration
• Licensing clarity per model family

🛡️ Security, Compatibility & Tooling

🔐 Security & Dependency Updates

• Dependency upgrades
• Internal security improvements
• Explicit workflow permissions added (#1021, #1023)

🐍 Python Version Support

• Dropped: Python 3.9
• Added: Python 3.13
• Supported: Python 3.10–3.13
• Updated CUDA wheel source to cu126

🛠️ Developer Tooling & CI/CD

• Expanded mypy type-checking coverage (#912, #931, #935, #951)
• Updated pre-commit hooks and general formatting
• CI uses CPU-only PyTorch for faster, more reliable builds (#974, #979)
• Updated pip install workflow (#1013)
• Added new Python 3.13 Docker images (#1014, #1019)

🧹 Bug Fixes & Stability Improvements

• Fixed multi-GPU behaviour with torch.compile (#923)
• Fixed DICOM reading issue (#934)
• Fixed annotation contour handling with holes (#956)
• Fixed consecutive annotation load bug (#927)
• Fixed SCCNN model issues (#970)
• Fixed MapDe dist_filter shape issue (#914)
• Improved notebook reliability on Colab (#1026–#1030)
• macOS TileServer issues resolved (#976)

🧭 Migration Guide for Users

🔄 Updating from 1.x to 2.0.0

Update calls: replace .predict() with .run()

# Old
results = segmentor.predict(imgs=[...], ioconfig=config)

# New
results = segmentor.run(images=[...], ioconfig=config)

Use patch_mode: replace mode="patch" with patch_mode=True and mode="tile" or "wsi" with patch_mode=False

# Old
results = segmentor.predict(imgs=[...], mode="patch", ioconfig=config)

# New
results = segmentor.run(images=[...], patch_mode=True, ioconfig=config)

# Old
results = segmentor.predict(imgs=[...], mode="wsi", ioconfig=config)

# New
results = segmentor.run(images=[...], patch_mode=False, ioconfig=config)

Use the new I/O configs

from tiatoolbox.models.engine.io_config import IOSegmentorConfig

config = IOSegmentorConfig(
    patch_input_shape=(256, 256),
    stride_shape=(240, 240),
    input_resolutions=[{"resolution": 0.25, "units": "mpp"}],
    save_resolution={"units": "baseline", "resolution": 1.0},
)

Specify the output format

results = segmentor.run(
    images=[...],
    ioconfig=ioconfig,
    output_type="zarr",  # or "dict", "annotationstore", "qupath"
    save_dir="outputs/",
)

Update imports

• tiatoolbox.typing → tiatoolbox.type_hints

Install requirements

• Python 3.10+ required
• On Windows: install OpenSlide via pip install openslide-bin

Full Changelog: TissueImageAnalytics/tiatoolbox@v1.6.0...v2.0.0