MONAI Framework

The foundational library providing domain-optimized data loaders, network architectures, and loss functions for medical imaging. It natively handles DICOM and NIfTI formats, supports 2D and 3D volumetric data, and includes specialized transforms like RandSpatialCropSamples and RandGaussianNoise for robust data augmentation. The CacheDataset mechanism intelligently caches transformed data to accelerate training loops, directly addressing the I/O bottleneck common in gigapixel pathology and high-resolution radiology workflows.

An intelligent, server-based annotation tool that uses active learning to reduce the manual effort required for creating ground-truth segmentation masks. It integrates with 3D Slicer and OHIF viewers, presenting radiologists with AI-generated initial segmentations for refinement. Key capabilities include:

• DeepGrow: An interactive segmentation method where users click foreground/background to iteratively refine boundaries.
• DeepEdit: Combines automated segmentation with user-provided scribbles for fine-grained control. This tight feedback loop accelerates the creation of high-quality labeled datasets for supervised learning.

A specification and toolkit for packaging trained models into production-ready MAPs (MONAI Application Packages). A MAP is a self-contained, containerized artifact that includes the model, pre/post-processing transforms, and an inference handler. It defines a strict App SDK contract, ensuring deterministic execution across development and deployment environments. This standard enables seamless integration with clinical PACS systems and DICOM routers, bridging the gap between research code and a validated, auditable clinical workflow.

A standardized format for packaging and sharing reproducible AI models. A bundle is a directory containing a PyTorch model checkpoint, a JSON/YAML configuration file defining the network, transforms, and training hyperparameters, and metadata. This approach solves the reproducibility crisis in medical AI research by ensuring that a published model can be loaded and executed with a single monai.bundle command, without manual reconstruction of the training pipeline. Bundles serve as the unit of exchange in the MONAI Model Zoo.

A public, community-driven repository of pre-trained MONAI Bundles for common medical imaging tasks. It hosts validated models for spleen CT segmentation, brain tumor MRI segmentation, and pathology nuclei classification, among others. Each model is accompanied by a model card documenting its training data, intended use, and performance metrics. This resource enables transfer learning and rapid prototyping, allowing teams to fine-tune state-of-the-art models on proprietary data rather than training from scratch.

A component designed for real-time AI inference on live video streams, such as endoscopy or surgical microscopy feeds. It implements an optimized pipeline that ingests frames, applies transforms, runs inference, and renders overlays with minimal latency. The architecture uses asynchronous processing and GPU-accelerated video I/O to maintain frame rates suitable for intraoperative guidance. This bridges the gap between offline batch processing and the demanding, low-latency requirements of interventional clinical settings.

The foundational PyTorch-based library providing domain-optimized data loaders, preprocessing transforms, and network architectures specifically designed for medical imaging. Unlike generic computer vision frameworks, MONAI Core handles DICOM and NIfTI formats natively, supports 3D volumetric data, and implements medical-specific augmentations such as elastic deformations and intensity normalization. The library's monai.transforms module offers a composable pipeline API that chains spatial, intensity, and IO operations with dictionary-based metadata tracking, ensuring that spatial transformations consistently update associated label maps and region-of-interest coordinates.

An intelligent interactive annotation tool that accelerates pathology and radiology ground-truth creation through AI-assisted labeling. MONAI Label integrates with 3D Slicer and OHIF viewers, providing a client-server architecture where server-side models propose segmentations that clinicians refine. Key capabilities include:

• Active learning loops that retrain models as new annotations are added
• Support for DeepEdit and DeepGrow interactive segmentation paradigms
• Built-in versioning for annotation provenance tracking This dramatically reduces the time required to create labeled datasets for whole slide image and volumetric CT analysis.

A specification and runtime for packaging, deploying, and orchestrating medical AI applications in clinical environments. MONAI Deploy defines the MONAI Application Package (MAP) standard—a containerized, self-describing format that bundles models, metadata, and execution requirements. The MONAI Deploy App SDK enables developers to build operator graphs that define deterministic processing pipelines, while the MONAI Deploy Informatics Gateway handles DICOM communication with PACS and VNA systems. This ensures that pathology classification models transition from research notebooks to production clinical workflows with guaranteed reproducibility.

A standardized format for packaging, sharing, and reproducing complete model workflows including preprocessing configurations, network definitions, and trained weights. A MONAI Bundle is a self-contained directory with a JSON/YAML metadata file that describes the model's purpose, inputs, outputs, and computational graph. Benefits for pathology teams include:

• Reproducible research with pinned dependency versions
• Model zoo integration for sharing pre-trained pathology foundation models
• Audit trail support for regulatory submissions Bundles enable seamless transfer of models between training, validation, and deployment environments without configuration drift.

A high-performance streaming data pipeline designed for ultra-low-latency inference on medical imaging streams. MONAI Stream addresses the bottleneck between data acquisition and AI processing by implementing zero-copy data transfer and GPU-accelerated decoding of compressed medical image formats. This is critical for real-time pathology applications such as live whole slide image analysis during scanning or intraoperative frozen section assessment, where sub-second inference latency directly impacts clinical decision-making. The architecture supports gRPC-based streaming protocols for integration with digital pathology scanners.

A federated learning framework that enables collaborative training of pathology models across multiple institutions without sharing sensitive patient data. MONAI FL implements secure aggregation protocols where only model weight updates—never raw images—leave each hospital's firewall. Key features include:

• Flexible topology supporting both client-server and peer-to-peer architectures
• Differential privacy integration for formal privacy guarantees
• Heterogeneous client support accommodating varying compute resources across sites This directly addresses the data silo problem in digital pathology, enabling robust model development on diverse, multi-institutional datasets.