Minor updates to ensure compatibility with HSDK v3.5.0

Min Python version is now 3.10
Use of NumPy can_cast with Python type is removed
Update the DICOM series to volume image to use pydicom apply_rescale directly
Tested Jupyter notebooks (using locally built SDK for monai-deploy package command)

Signed-off-by: M Q <[email protected]>

Author: MMelQin

.github/workflows/pr.yml

@@ -12,12 +12,15 @@ on:
 jobs:
   test:
     runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: [ "3.10", "3.11", "3.12", "3.13" ]
     steps:
       - uses: actions/checkout@v2
-      - name: Set up Python 3.9
+      - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@v2
         with:
-          python-version: "3.9"
+          python-version: ${{ matrix.python-version }}
       - name: Setup Dev Environment
         run: |
           pip install virtualenv

README.md

@@ -41,7 +41,7 @@ pip install monai-deploy-app-sdk
 - This SDK depends on [NVIDIA Holoscan SDK](https://pypi.org/project/holoscan/) for its core implementation as well as its CLI, hence inherits its prerequisites, e.g. Ubuntu 22.04 with glibc 2.35 on X86-64 and NVIDIA dGPU drivers version 535 or above.
 - [CUDA 12.2](https://developer.nvidia.com/cuda-12-2-0-download-archive) or above is required along with a supported NVIDIA GPU with at least 8GB of video RAM.
 - If inference is not used in an example application and a GPU is not installed, at least [CUDA 12 runtime](https://pypi.org/project/nvidia-cuda-runtime-cu12/) is required, as this is one of the requirements of Holoscan SDK. In addition, the `LIB_LIBRARY_PATH` must be set to include the installed shared library, e.g. in a Python 3.10 env, ```export LD_LIBRARY_PATH=`pwd`/.venv/lib/python3.10/site-packages/nvidia/cuda_runtime/lib:$LD_LIBRARY_PATH```
-- Python: 3.9 to 3.12
+- Python: 3.10 to 3.13
 
 ## Getting Started
 

docs/source/getting_started/tutorials/mednist_app.md

@@ -5,9 +5,9 @@ This tutorial demos the process of packaging up a trained model using MONAI Depl
 ## Setup
 
 ```bash
-# Create a virtual environment with Python 3.9.
+# Create a virtual environment with Python 3.10.
 # Skip if you are already in a virtual environment.
-conda create -n mednist python=3.9 pytorch jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
+conda create -n mednist python=3.10 pytorch jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
 conda activate mednist
 
 # Launch JupyterLab if you want to work on Jupyter Notebook

docs/source/getting_started/tutorials/monai_bundle_app.md

@@ -5,9 +5,9 @@ This tutorial shows how to create an organ segmentation application for a PyTorc
 ## Setup
 
 ```bash
-# Create a virtual environment with Python 3.9.
+# Create a virtual environment with Python 3.10.
 # Skip if you are already in a virtual environment.
-conda create -n monai python=3.9 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
+conda create -n monai python=3.10 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
 conda activate monai
 
 # Launch JupyterLab if you want to work on Jupyter Notebook

docs/source/getting_started/tutorials/multi_model_app.md

@@ -7,9 +7,9 @@ The models used in this example are trained with MONAI, and are packaged in the
 ## Setup
 
 ```bash
-# Create a virtual environment with Python 3.9.
+# Create a virtual environment with Python 3.10.
 # Skip if you are already in a virtual environment.
-conda create -n monai python=3.9 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
+conda create -n monai python=3.10 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
 conda activate monai
 
 # Launch JupyterLab if you want to work on Jupyter Notebook

docs/source/getting_started/tutorials/segmentation_app.md

@@ -7,9 +7,9 @@ Please note that the following steps are for demonstration purpose. The code pul
 ## Setup
 
 ```bash
-# Create a virtual environment with Python 3.9.
+# Create a virtual environment with Python 3.10.
 # Skip if you are already in a virtual environment.
-conda create -n monai python=3.9 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
+conda create -n monai python=3.10 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
 conda activate monai
 
 # Launch JupyterLab if you want to work on Jupyter Notebook

docs/source/getting_started/tutorials/segmentation_clara-viz_app.md

@@ -5,9 +5,9 @@ This tutorial shows how to create an organ segmentation application for a PyTorc
 ## Setup
 
 ```bash
-# Create a virtual environment with Python 3.9.
+# Create a virtual environment with Python 3.10.
 # Skip if you are already in a virtual environment.
-conda create -n monai python=3.9 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
+conda create -n monai python=3.10 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
 conda activate monai
 
 # Launch JupyterLab if you want to work on Jupyter Notebook

docs/source/getting_started/tutorials/simple_app.md

@@ -5,9 +5,9 @@ This tutorial shows how a simple image processing application can be created wit
 ## Setup
 
 ```bash
-# Create a virtual environment with Python 3.9.
+# Create a virtual environment with Python 3.10.
 # Skip if you are already in a virtual environment.
-conda create -n monai python=3.9 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
+conda create -n monai python=3.10 pytorch torchvision jupyterlab cudatoolkit=12.2 -c pytorch -c conda-forge
 conda activate monai
 
 # Launch JupyterLab if you want to work on Jupyter Notebook

monai/deploy/operators/dicom_series_to_volume_operator.py

@@ -16,6 +16,10 @@
 
 import numpy as np
 
+from monai.deploy.utils.importutil import optional_import
+
+apply_rescale, _ = optional_import("pydicom.pixels", name="apply_rescale")
+
 from monai.deploy.core import ConditionType, Fragment, Operator, OperatorSpec
 from monai.deploy.core.domain.dicom_series_selection import StudySelectedSeries
 from monai.deploy.core.domain.image import Image
@@ -112,10 +116,13 @@ def generate_voxel_data(self, series):
         # with the NumPy array returned from the ITK GetArrayViewFromImage on the image
         # loaded from the same DICOM series.
         vol_data = np.stack([s.get_pixel_array() for s in slices], axis=0)
-        # The above get_pixel_array() already considers the PixelRepresentation attribute,
-        # 0 is unsigned int, 1 is signed int
-        if slices[0][0x0028, 0x0103].value == 0:
-            vol_data = vol_data.astype(np.uint16)
+
+        # Use pydicom utility to apply a modality lookup table or rescale operator to the pixel array.
+        # The pydicom Dataset is required which can be obtained from the first slice's native SOP instance.
+        # If Modality LUT is present the return array is of np.uint8 or np.uint16, and if Rescale
+        # Intercept and Rescale Slope are present, np.float64.
+        # If the pixel array is already in the correct type, the return array is the same as the input array.
+        vol_data = apply_rescale(vol_data, slices[0].get_native_sop_instance())
 
         # For now we support monochrome image only, for which DICOM Photometric Interpretation
         # (0028,0004) has defined terms, MONOCHROME1 and MONOCHROME2, with the former being:
@@ -146,50 +153,20 @@ def generate_voxel_data(self, series):
                     f"Cannot process pixel data with Photometric Interpretation of {photometric_interpretation}."
                 )
 
-        # Rescale Intercept and Slope attributes might be missing, but safe to assume defaults.
-        try:
-            intercept = slices[0][0x0028, 0x1052].value
-        except KeyError:
-            intercept = 0
-
-        try:
-            slope = slices[0][0x0028, 0x1053].value
-        except KeyError:
-            slope = 1
-
-        # check if vol_data, intercept, and slope can be cast to uint16 without data loss
-        if (
-            np.can_cast(vol_data, np.uint16, casting="safe")
-            and np.can_cast(intercept, np.uint16, casting="safe")
-            and np.can_cast(slope, np.uint16, casting="safe")
-        ):
+        # NumPy's np.can_cast function, as of version 2.0, no longer supports Python scalars directly and
+        # does not apply value-based logic for 0-D arrays and NumPy scalars.
+        # The following can_cast calls are expecting the array is already of the correct type.
+        if vol_data.dtype == np.uint8:
+            logging.info("Rescaled pixel data is of type uint8.")
+        elif np.can_cast(vol_data, np.uint16, casting="safe"):
             logging.info("Casting to uint16")
-            vol_data = np.array(vol_data, dtype=np.uint16)
-            intercept = np.uint16(intercept)
-            slope = np.uint16(slope)
-        elif (
-            np.can_cast(vol_data, np.float32, casting="safe")
-            and np.can_cast(intercept, np.float32, casting="safe")
-            and np.can_cast(slope, np.float32, casting="safe")
-        ):
+            vol_data = vol_data.astype(dtype=np.uint16, casting="safe")
+        elif np.can_cast(vol_data, np.float32, casting="safe"):
             logging.info("Casting to float32")
-            vol_data = np.array(vol_data, dtype=np.float32)
-            intercept = np.float32(intercept)
-            slope = np.float32(slope)
-        elif (
-            np.can_cast(vol_data, np.float64, casting="safe")
-            and np.can_cast(intercept, np.float64, casting="safe")
-            and np.can_cast(slope, np.float64, casting="safe")
-        ):
-            logging.info("Casting to float64")
-            vol_data = np.array(vol_data, dtype=np.float64)
-            intercept = np.float64(intercept)
-            slope = np.float64(slope)
-
-        if slope != 1:
-            vol_data = slope * vol_data
-
-        vol_data += intercept
+            vol_data = vol_data.astype(dtype=np.float32, casting="safe")
+        else:
+            logging.info("Rescaled pixel data remains as of type float64.")
+
         return vol_data
 
     def create_volumetric_image(self, vox_data, metadata):