Add GPU-accelerated dcm image decoding

docs/source/data.rst

@@ -159,6 +159,21 @@ PILReader
 .. autoclass:: PILReader
   :members:
 
+PydicomReader
+~~~~~~~~~~~~~
+.. autoclass:: PydicomReader
+  :members:
+
+NvImgCodecPydicomReader
+~~~~~~~~~~~~~~~~~~~~~~~
+GPU-accelerated DICOM reader built on :py:class:`PydicomReader` and the nvImageCodec pydicom decoder plugin.
+The ``to_gpu`` init argument is accepted for API compatibility but is always ignored so that GPU decompression
+is not bypassed by GPU direct loading. The optional dependency ``nvimgcodec`` needs to be installed for this
+feature (``pip install 'monai[nvimgcodec]'``).
+
+.. autoclass:: NvImgCodecPydicomReader
+  :members:
+
 NrrdReader
 ~~~~~~~~~~
 .. autoclass:: NrrdReader

docs/source/installation.md

@@ -77,13 +77,17 @@ pip install monai
 MONAI supports the extras syntax such as `pip install 'monai[nibabel]'`. The options are
 
 ```text
-clearml, cucim, cupy, einops, fire, gdown, h5py, huggingface_hub, hyena, ignite, imagecodecs, itk, jsonschema, lmdb, lpips, matplotlib, metrics_reloaded, mlflow, nibabel, nni, onnx, openslide, optuna, pandas, pillow, polygraphy, psutil, pyamg, pybind11, pydicom, pynrrd, pynvml, pyyaml, requests, segment_anything, scipy, skimage, tensorboard, tensorboardX, tifffile, torchio, torchvision, tqdm, transformers, zarr
+clearml, cucim, cupy, einops, fire, gdown, h5py, huggingface_hub, hyena, ignite, imagecodecs, itk, jsonschema, lmdb, lpips, matplotlib, metrics_reloaded, mlflow, nibabel, nni, nvimgcodec, onnx, openslide, optuna, pandas, pillow, polygraphy, psutil, pyamg, pybind11, pydicom, pynrrd, pynvml, pyyaml, requests, segment_anything, scipy, skimage, tensorboard, tensorboardX, tifffile, torchio, torchvision, tqdm, transformers, zarr
 ```
 
-which correspond to the packages: `clearml`, `cucim` (`cucim-cu12` or `cucim-cu13`), `cupy-cuda13x`, `einops`, `fire`, `gdown`, `h5py`, `huggingface_hub`, `nvsubquadratic`, `omegaconf`, `pytorch-ignite`, `imagecodecs`, `itk`, `jsonschema`, `lmdb`, `lpips`, `matplotlib`, `MetricsReloaded`, `mlflow`, `nibabel`, `nni`, `filelock`, `onnx`, `onnxruntime`, `onnx_graphsurgeon`, `onnxscript`, `openslide-python`, `openslide-bin`, `optuna`, `pandas`, `pillow`, `polygraphy`, `psutil`, `pyamg`, `pybind11`, `pydicom`, `pynrrd`, `nvidia-ml-py`, `pyyaml`, `requests`, `segment_anything`, `scipy`, `scikit-image`, `tensorboard`, `tensorboardX`, `tifffile`, `torchio`, `torchvision`, `tqdm`, `transformers`, `zarr`.
+which correspond to the packages: `clearml`, `cucim` (`cucim-cu12` or `cucim-cu13`), `cupy-cuda13x`, `einops`, `fire`, `gdown`, `h5py`, `huggingface_hub`, `nvsubquadratic`, `omegaconf`, `pytorch-ignite`, `imagecodecs`, `itk`, `jsonschema`, `lmdb`, `lpips`, `matplotlib`, `MetricsReloaded`, `mlflow`, `nibabel`, `nni`, `filelock`, `nvidia-nvimgcodec-cu13`, `onnx`, `onnxruntime`, `onnx_graphsurgeon`, `onnxscript`, `openslide-python`, `openslide-bin`, `optuna`, `pandas`, `pillow`, `polygraphy`, `psutil`, `pyamg`, `pybind11`, `pydicom`, `pynrrd`, `nvidia-ml-py`, `pyyaml`, `requests`, `segment_anything`, `scipy`, `scikit-image`, `tensorboard`, `tensorboardX`, `tifffile`, `torchio`, `torchvision`, `tqdm`, `transformers`, `zarr`.
 
-Almost all of these can be installed together with the `all` option. For development on MONAI, this should be accompanied by `testing` which will install the testing static checking packages. Cupy is omitted from `all` since the choice between
-Cuda 12 and 13 versions of the library can't be resolved when installing and must be manually installed.
+Almost all of these can be installed together with the `all` option. For development on MONAI, this should be accompanied by `testing` which will install the testing static checking packages. Cupy and `nvimgcodec` are omitted from `all` since the choice between
+Cuda 12 and 13 versions of the libraries can't be resolved when installing and must be manually installed.
+
+The `nvimgcodec` extra installs GPU-accelerated DICOM decoding for `NvImgCodecPydicomReader`
+(`pip install 'monai[nvimgcodec]'`). It is Linux-only in the extra definition; CUDA 13 is the
+default. CUDA 12 users should install matching `cupy-cuda12x` and `nvidia-nvimgcodec-cu12` wheels.
 
 The `hyena` extra pulls in [`nvsubquadratic`](https://github.com/NVIDIA-BioNeMo/nvSubquadratic),
 required by `HyenaNDUNETR` / `HyenaMixer` / `HyenaTransformerBlock` (subquadratic

monai/data/__init__.py

@@ -50,7 +50,16 @@
 from .folder_layout import FolderLayout, FolderLayoutBase
 from .grid_dataset import GridPatchDataset, PatchDataset, PatchIter, PatchIterd
 from .image_dataset import ImageDataset
-from .image_reader import ImageReader, ITKReader, NibabelReader, NrrdReader, NumpyReader, PILReader, PydicomReader
+from .image_reader import (
+    ImageReader,
+    ITKReader,
+    NibabelReader,
+    NrrdReader,
+    NumpyReader,
+    NvImgCodecPydicomReader,
+    PILReader,
+    PydicomReader,
+)
 from .image_writer import (
     SUPPORTED_WRITERS,
     ImageWriter,

monai/data/image_reader.py

@@ -64,7 +64,22 @@
 else:
     NdarrayOrCupy: TypeAlias = Any
 
-__all__ = ["ImageReader", "ITKReader", "NibabelReader", "NumpyReader", "PILReader", "PydicomReader", "NrrdReader"]
+__all__ = [
+    "ImageReader",
+    "ITKReader",
+    "NibabelReader",
+    "NumpyReader",
+    "PILReader",
+    "PydicomReader",
+    "NvImgCodecPydicomReader",
+    "NrrdReader",
+    "DICOM_READER_ENV_MAP",
+    "get_preferred_dicom_reader_key",
+    "is_dicom_path",
+]
+
+# Maps ``MONAI_DICOM_READER`` env values to keys in :py:data:`monai.transforms.io.array.SUPPORTED_READERS`.
+DICOM_READER_ENV_MAP = {"itk": "itkreader", "pydicom": "pydicomreader", "nvimgcodec": "nvimgcodecpydicomreader"}
 
 
 class ImageReader(ABC):
@@ -1027,6 +1042,141 @@ def _get_array_data(self, img, filename):
         return data
 
 
+def is_dicom_path(filename: Sequence[PathLike] | PathLike) -> bool:
+    """
+    Return ``True`` if ``filename`` refers to a DICOM file or a directory that may contain a DICOM series.
+    """
+    for name in ensure_tuple(filename):
+        name = f"{name}"
+        path = Path(name)
+        if path.is_dir():
+            return True
+        if path.suffix.lower() == ".dcm":
+            return True
+        if has_pydicom:
+            try:
+                if pydicom.misc.is_dicom(name):
+                    return True
+            except Exception:
+                pass
+    return False
+
+
+def get_preferred_dicom_reader_key() -> str:
+    """
+    Return the :py:class:`~monai.transforms.LoadImage` registration key for the preferred DICOM reader.
+
+    Controlled by the ``MONAI_DICOM_READER`` environment variable. Supported values are
+    ``itk``, ``pydicom``, and ``nvimgcodec``. Returns an empty string when the variable is
+    unset or set to an unsupported value (in which case :py:data:`~monai.transforms.io.array.SUPPORTED_READERS`
+    dict order is used unchanged).
+    """
+    pref = os.environ.get("MONAI_DICOM_READER")
+    if pref is None:
+        return ""
+    pref = pref.lower()
+    if pref not in DICOM_READER_ENV_MAP:
+        warnings.warn(f"Unknown MONAI_DICOM_READER='{pref}', ignoring preference.")
+        return ""
+    return DICOM_READER_ENV_MAP[pref]
+
+
+@require_pkg(pkg_name="pydicom")
+class NvImgCodecPydicomReader(PydicomReader):
+    """
+    Load DICOM images using Pydicom with GPU-accelerated decompression via nvImageCodec.
+
+    This reader extends :py:class:`PydicomReader` and registers the nvImageCodec pydicom
+    decoder plugin on initialization. The plugin accelerates decoding of compressed pixel data
+    for JPEG, JPEG 2000, and HTJ2K transfer syntaxes when CUDA, CuPy and ``nvidia-nvimgcodec`` are available.
+
+    If nvImageCodec is not available, a warning is issued and the reader falls back to the
+    default pydicom decoders (same behavior as :py:class:`PydicomReader`).
+
+    Requires the optional extra ``nvimgcodec`` (``pip install 'monai[nvimgcodec]'``), which installs ``pydicom``,
+    CuPy, and ``nvidia-nvimgcodec-cu13`` on Linux. GPU decompression uses
+    ``nvidia.nvimgcodec.tools.dicom.pydicom_plugin`` from the nvImageCodec package. CUDA 13 is strongly
+    recommended because the nvJPEG library has addressed a known issue with JPEG lossless decoding in
+    CUDA 13.2.0+. For CUDA 12, install matching ``cupy-cuda12x`` and ``nvidia-nvimgcodec-cu12`` wheels instead.
+
+    Set environment variable ``MONAI_DICOM_READER=nvimgcodec`` to use this reader by default
+    with :py:class:`monai.transforms.LoadImage` without explicit configuration.
+
+    Note:
+        GPU direct loading bypasses Pydicom pixel data interpretation mechanism hence disables GPU decompression
+        via Pydicom decoder plugin that is used by this reader. So, GPU direct loading (``to_gpu=True``)
+        cannot be supported by this reader. The ``to_gpu`` init argument is accepted for API compatibility
+        with :py:class:`PydicomReader` but is always ignored so that GPU-accelerated decompression via nvImageCodec
+        is not bypassed.
+
+        Also noted is that the current implementation of GPU direct loading has a serious flaw as it simply loads
+        the raw bytes of pixel data into GPU memory and parses them into integers without any required processing,
+        e.g. applying rescale slope and intercept, `PhotometricInterpretation`, etc., and not processing compressed
+        pixel data. As such, the resulting data array will not represent the original pixel data faithfully except for
+        the simplest case of uncompressed pixel data.
+
+        This reader only declares ``@require_pkg(pkg_name="pydicom")`` so that :py:class:`monai.transforms.LoadImage`
+        can register it without hard-failing when GPU dependencies are missing. ``pydicom`` is required to construct
+        the reader; nvimgcodec, CUDA, and CuPy availability is checked at runtime with a warning issued and fallback to
+        default pydicom decoders if missing.
+
+    Args:
+        channel_dim: the channel dimension of the input image, default is None.
+            This is used to set original_channel_dim in the metadata, EnsureChannelFirstD reads this field.
+            If None, `original_channel_dim` will be either `no_channel` or `-1`.
+        affine_lps_to_ras: whether to convert the affine matrix from "LPS" to "RAS". Defaults to ``True``.
+        swap_ij: whether to swap the first two spatial axes. Default to ``True``.
+        prune_metadata: whether to prune the saved information in metadata. Default to ``True``.
+        label_dict: label of the dicom data for segmentation loading.
+        fname_regex: a regular expression to match file names when the input is a folder.
+        to_gpu: accepted for API compatibility with :py:class:`PydicomReader` but always ignored (always ``False``).
+        kwargs: additional args for `pydicom.dcmread` API.
+    """
+
+    def __init__(
+        self,
+        channel_dim: str | int | None = None,
+        affine_lps_to_ras: bool = True,
+        swap_ij: bool = True,
+        prune_metadata: bool = True,
+        label_dict: dict | None = None,
+        fname_regex: str = "",
+        to_gpu: bool = False,
+        **kwargs,
+    ):
+        if to_gpu:
+            warnings.warn(
+                "NvImgCodecPydicomReader ignores to_gpu=True; GPU direct loading is disabled to preserve "
+                "GPU-accelerated decompression."
+            )
+        super().__init__(
+            channel_dim=channel_dim,
+            affine_lps_to_ras=affine_lps_to_ras,
+            swap_ij=swap_ij,
+            prune_metadata=prune_metadata,
+            label_dict=label_dict,
+            fname_regex=fname_regex,
+            to_gpu=False,
+            **kwargs,
+        )
+        from monai.data.nvimgcodec_pydicom_plugin import is_nvimgcodec_available, register_as_decoder_plugin
+
+        self._nvimgcodec_available = is_nvimgcodec_available()
+        if not register_as_decoder_plugin():
+            warnings.warn(
+                "NvImgCodecPydicomReader: nvImageCodec decoder plugin did not register successfully. "
+                "Falling back to default pydicom decoders."
+            )
+
+    def verify_suffix(self, filename: Sequence[PathLike] | PathLike) -> bool:
+        """
+        Verify whether the specified file or files are DICOM and nvImageCodec is available.
+        """
+        if not has_pydicom or not self._nvimgcodec_available:
+            return False
+        return is_dicom_path(filename)
+
+
 @require_pkg(pkg_name="nibabel")
 class NibabelReader(ImageReader):
     """
@@ -1235,7 +1385,7 @@ def _get_array_data(self, img, filename):
             with kvikio.CuFile(filename, "r") as f:
                 f.read(image)
             if filename.endswith(".nii.gz"):
-                # for compressed data, have to tansfer to CPU to decompress
+                # for compressed data, have to transfer to CPU to decompress
                 # and then transfer back to GPU. It is not efficient compared to .nii file
                 # and may be slower than CPU loading in some cases.
                 warnings.warn("Loading compressed NIfTI file into GPU may not be efficient.")

monai/data/nvimgcodec_pydicom_plugin.py

@@ -0,0 +1,80 @@
+# Copyright (c) MONAI Consortium
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+MONAI integration helpers for the nvImageCodec pydicom decoder plugin.
+
+The decoder implementation lives in ``nvidia.nvimgcodec.tools.dicom.pydicom_plugin``
+(shipped with ``nvidia-nvimgcodec-cuXX``). This module provides MONAI-facing helpers
+and stable aliases for registration and availability checks.
+"""
+
+from __future__ import annotations
+
+import logging
+
+from monai.utils import optional_import
+
+cp, has_cp = optional_import("cupy")
+pydicom_plugin, has_pydicom_plugin = optional_import("nvidia.nvimgcodec.tools.dicom.pydicom_plugin")
+
+_logger = logging.getLogger(__name__)
+
+if has_pydicom_plugin:
+    DECODER_DEPENDENCIES = pydicom_plugin.DECODER_DEPENDENCIES
+    NVIMGCODEC_MIN_VERSION = pydicom_plugin.NVIMGCODEC_MIN_VERSION
+    NVIMGCODEC_MIN_VERSION_TUPLE = pydicom_plugin.NVIMGCODEC_MIN_VERSION_TUPLE
+    NVIMGCODEC_PLUGIN_LABEL = pydicom_plugin.NVIMGCODEC_PLUGIN_LABEL
+    SUPPORTED_DECODER_CLASSES = pydicom_plugin.SUPPORTED_DECODER_CLASSES
+    SUPPORTED_TRANSFER_SYNTAXES = pydicom_plugin.SUPPORTED_TRANSFER_SYNTAXES
+    is_available = pydicom_plugin.is_available
+else:  # pragma: no cover - optional dependency not installed
+    DECODER_DEPENDENCIES = {}
+    NVIMGCODEC_MIN_VERSION = "0.8.0"
+    NVIMGCODEC_MIN_VERSION_TUPLE = (0, 8, 0)
+    NVIMGCODEC_PLUGIN_LABEL = "0.8.0+nvimgcodec"
+    SUPPORTED_DECODER_CLASSES = []
+    SUPPORTED_TRANSFER_SYNTAXES = []
+
+    def is_available(uid) -> bool:  # type: ignore[no-redef]
+        return False
+
+
+def is_nvimgcodec_available() -> bool:
+    """Return ``True`` if nvImageCodec with CUDA support is available."""
+    if not has_pydicom_plugin or getattr(pydicom_plugin, "nvimgcodec", None) is None or not has_cp:
+        _logger.debug("nvimgcodec pydicom plugin, nvimgcodec module, or CuPy missing.")
+        return False
+    try:
+        if not cp.cuda.is_available():
+            _logger.debug("CUDA device not found.")
+            return False
+    except Exception as exc:  # pragma: no cover - environment specific
+        _logger.debug(f"CUDA availability check failed: {exc}")
+        return False
+    return True
+
+
+def register_as_decoder_plugin(module_path: str | None = None) -> bool:
+    """Register the nvImageCodec pydicom decoder plugin."""
+    if not is_nvimgcodec_available():
+        _logger.warning("nvImageCodec is not available; skipping pydicom decoder plugin registration.")
+        return False
+    if not has_pydicom_plugin:
+        return False
+    return bool(pydicom_plugin.register(module_path))
+
+
+def unregister_as_decoder_plugin() -> bool:
+    """Unregister the nvImageCodec pydicom decoder plugin."""
+    if not has_pydicom_plugin:
+        return False
+    return bool(pydicom_plugin.unregister())

monai/transforms/io/array.py

@@ -36,8 +36,10 @@
     NibabelReader,
     NrrdReader,
     NumpyReader,
+    NvImgCodecPydicomReader,
     PILReader,
     PydicomReader,
+    get_preferred_dicom_reader_key,
 )
 from monai.data.meta_tensor import MetaTensor
 from monai.data.utils import is_no_channel
@@ -59,10 +61,15 @@
 nrrd, _ = optional_import("nrrd")
 FileLock, has_filelock = optional_import("filelock", name="FileLock")
 
-__all__ = ["LoadImage", "SaveImage", "SUPPORTED_READERS"]
+__all__ = ["LoadImage", "SaveImage", "SUPPORTED_READERS", "get_default_reader_registration_order"]
 
+# Default readers for :py:class:`LoadImage`. Dict insertion order is the registration order
+# (auto-select tries registered readers from last to first). Users may add custom readers here.
+# DICOM readers are listed first so that, by default, ``itkreader`` is tried before the other
+# DICOM readers; ``MONAI_DICOM_READER`` can promote a preferred DICOM reader to last (tried first).
 SUPPORTED_READERS = {
     "pydicomreader": PydicomReader,
+    "nvimgcodecpydicomreader": NvImgCodecPydicomReader,
     "itkreader": ITKReader,
     "nrrdreader": NrrdReader,
     "numpyreader": NumpyReader,
@@ -71,6 +78,22 @@
 }
 
 
+def get_default_reader_registration_order() -> list[str]:
+    """
+    Return the default reader registration order for :py:class:`LoadImage`.
+
+    Uses :py:data:`SUPPORTED_READERS` insertion order so user-added entries are included.
+    If ``MONAI_DICOM_READER`` resolves to a non-empty preferred key present in
+    ``SUPPORTED_READERS``, that key is moved to the end of the list so auto-selection
+    tries it first.
+    """
+    order = list(SUPPORTED_READERS)
+    preferred = get_preferred_dicom_reader_key()
+    if preferred and preferred in order:
+        order = [key for key in order if key != preferred] + [preferred]
+    return order
+
+
 def switch_endianness(data, new="<"):
     """
     Convert the input `data` endianness to `new`.
@@ -116,7 +139,10 @@ class LoadImage(Transform):
         - User-specified reader in the constructor of `LoadImage`.
         - Readers from the last to the first in the registered list.
         - Current default readers: (nii, nii.gz -> NibabelReader), (png, jpg, bmp -> PILReader),
-          (npz, npy -> NumpyReader), (nrrd -> NrrdReader), (DICOM file -> ITKReader).
+          (npz, npy -> NumpyReader), (nrrd -> NrrdReader),
+          (DICOM file -> ITKReader first among DICOM readers by default).
+        - Optionally set ``MONAI_DICOM_READER`` to ``itk``, ``pydicom``, or ``nvimgcodec``
+          (GPU-accelerated decoding; requires ``pip install 'monai[nvimgcodec]'``) to try that DICOM reader first.
 
     Please note that for png, jpg, bmp, and other 2D formats, readers by default swap axis 0 and 1 after
     loading the array with ``reverse_indexing`` set to ``True`` because the spatial axes definition
@@ -185,7 +211,7 @@ def __init__(
         self.expanduser = expanduser
 
         self.readers: list[ImageReader] = []
-        for r in SUPPORTED_READERS:  # set predefined readers as default
+        for r in get_default_reader_registration_order():  # set predefined readers as default
             try:
                 self.register(SUPPORTED_READERS[r](*args, **kwargs))
             except OptionalImportError: