Supported features and limitations

2025-09-26

2 min read time

Applies to Linux

This topic discusses the features and limitations for MONAI 1.0.0 for AMD ROCm.

Features

Here are the MONAI for AMD ROCm features:

• Deep learning inference

  • Accelerated inference for MONAI models using AMD ROCm and HIP backends.

  • Supports common MONAI model architectures for segmentation, classification, and registration, along with advanced AI capabilities such as generative models, federated learning, and AutoML.

• Seamless integration with hipCIM

  • Supports hipCIM for pre-processing and post-processing of medical images.

  • Efficiently handles whole-slide imaging (WSI) data through GPU-optimized implementations of color augmentation, spatial transformations, and intensity scaling.

  • Provides accelerated transforms, morphological operations, and data augmentation that outperform CPU-only pipelines, particularly in workflows such as whole-slide imaging, patch extraction, and real-time augmentation.

• GPU acceleration

  • Leverages AMD Instinct GPUs for high-throughput inference.

  • Delivers optimized memory and compute performance for large-scale medical datasets.

• Extensibility

  • Compatible with MONAI’s modular design.

  • Supports Python APIs and plugin-based extensions.

• Interoperability

  • Works with PyTorch for AMD ROCm.

  • Compatible with CuPy and other GPU-accelerated libraries.

• Optimized for 3D medical imaging

  • Supports CT, MRI, Ultrasound, and other volumetric modalities with domain-specific optimizations.

• Prebuilt training pipelines optimized for AMD Instinct GPUs

  • Supports segmentation, classification, and detection tasks, reducing setup overhead.

• Model Zoo with pretrained models

  • Provides access to a wide collection of pretrained models from the MONAI Model Zoo, ready for fine-tuning on custom datasets.

  • Facilitates utilizing the MONAI Bundle format to easily get started on building workflows or integrating new models into your projects.

  For more information on Model Zoo, see MONAI Model Zoo.

Limitations

• MONAI for AMD ROCm only supports features from amd-cupy later than 13.5.1 and hipCIM 1.0.00 and later.

• There is no support for:

  • GPU direct storage (KvikIO, cuFile).

  • rocTX tracing.

• No support for Python earlier than 3.10 and PyTorch earlier than 1.13.1.

• Deprecated transforms such as AddChannel, AsChannelFirst, and others.

• There might not be first-class support for some advanced or rare image file formats and non-NIfTI/DICOM derivatives.

• No support for legacy neural network architectures such as deprecated versions of DynUnet and old TorchVision wrappers.

• Automatic installation of optional dependencies is not available. Some features require explicit installation.