diff --git a/.github/workflows/build-test-package.yml b/.github/workflows/build-test-package.yml
index b0b6759d8..fd3769af9 100644
--- a/.github/workflows/build-test-package.yml
+++ b/.github/workflows/build-test-package.yml
@@ -227,7 +227,7 @@ jobs:
         for tarball in "-manylinux_2_28" "-manylinux2014"; do
            rm -rf ITKPythonPackage
            export TARBALL_SPECIALIZATION=${tarball}
-           ./dockcross-manylinux-download-cache-and-build-module-wheels.sh -c "-DCUDAToolkit_ROOT=/usr/lib64/cuda116 -DCMAKE_CUDA_COMPILER=/usr/lib64/cuda116/bin/nvcc" -x "libcuda.so;libcuda.so.1;libcudart.so;libcudart.so.11.0;libcublas.so;libcublas.so.11;libcublasLt.so;libcublasLt.so.11;libcufft.so;libcufft.so.10" cp${{ matrix.python-version }}
+           ./dockcross-manylinux-download-cache-and-build-module-wheels.sh -c "-DCUDAToolkit_ROOT=/usr/lib64/cuda116 -DCMAKE_CUDA_COMPILER=/usr/lib64/cuda116/bin/nvcc -DRTK_CUDA_VERSION=11.6" -x "libcuda.so;libcuda.so.1;libcudart.so;libcudart.so.11.0;libcublas.so;libcublas.so.11;libcublasLt.so;libcublasLt.so.11;libcufft.so;libcufft.so.10" cp${{ matrix.python-version }}
         done
 
 
@@ -361,7 +361,7 @@ jobs:
         set PATH=C:\P\grep;%PATH%
         set CC=cl.exe
         set CXX=cl.exe
-        C:\Python3${{ matrix.python-version-minor }}-x64\python.exe C:\P\IPP\scripts\windows_build_module_wheels.py --py-envs "3${{ matrix.python-version-minor }}-x64" --no-cleanup --lib-paths "C:/Program\ Files/NVIDIA\ GPU\ Computing\ Toolkit/CUDA/v11.6/bin" --exclude-libs "nvcuda.dll;concrt140.dll;cublas64_11.dll;cublasLt64_11.dll;cudart64_110.dll;cufft64_10.dll"
+        C:\Python3${{ matrix.python-version-minor }}-x64\python.exe C:\P\IPP\scripts\windows_build_module_wheels.py --py-envs "3${{ matrix.python-version-minor }}-x64" --no-cleanup --lib-paths "C:/Program\ Files/NVIDIA\ GPU\ Computing\ Toolkit/CUDA/v11.6/bin" --exclude-libs "nvcuda.dll;concrt140.dll;cublas64_11.dll;cublasLt64_11.dll;cudart64_110.dll;cufft64_10.dll" -- "-DRTK_CUDA_VERSION=11.6"
 
     - name: Publish Python package as GitHub Artifact
       uses: actions/upload-artifact@v1
diff --git a/INSTALLATION.md b/INSTALLATION.md
index 2b6f95b0e..a695bac4f 100644
--- a/INSTALLATION.md
+++ b/INSTALLATION.md
@@ -9,6 +9,7 @@ RTK is a module of [ITK](https://www.itk.org), the Insight Toolkit. Follow the i
 * `Module_RTK_GIT_TAG`: Git tag for the RTK download. By default, the RTK version which is downloaded and compiled is the one given in the [RTK.remote.cmake](https://github.com/InsightSoftwareConsortium/ITK/blob/master/Modules/Remote/RTK.remote.cmake). Change this option to build another version. For example, you can change it to `master` to build the latest RTK version. RTK is only maintained to be backward compatible with the latest ITK release and ITK master branch.
 * `RTK_BUILD_APPLICATIONS`: Activates the compilation of RTK's command line tools. Although RTK is mainly a toolkit, we also provide several command line tools for doing most of the available processing. These command line tools use [gengetopt](https://www.gnu.org/software/gengetopt/gengetopt.html). Several examples are available on the [Applications](https://wiki.openrtk.org/index.php/RTK_wiki_help#Applications) section of the [wiki](http://wikiopenrtk.org).
 * `RTK_USE_CUDA`: Activates CUDA computation. Default is `ON` if CMake has automatically found the CUDA package and a CUDA-compatible GPU, and `OFF` otherwise.
+* `RTK_CUDA_VERSION`: Specifies an exact version of the CUDA toolkit which must be used. If unspecified, RTK only checks if the found version is recent enough.
 * `RTK_CUDA_PROJECTIONS_SLAB_SIZE`: Set the number of projections processed at once in CUDA processing. Default is 16.
 * `RTK_PROBE_EACH_FILTER`: Activates the timing, CPU and CUDA memory consumption of each filter. Defaults is `OFF`. When activated, each filter processing is probed and a summary can be displayed. All command line applications display the result with `--verbose`.
 
diff --git a/itk-module-init.cmake b/itk-module-init.cmake
index 9a7e0785d..ee9172caa 100644
--- a/itk-module-init.cmake
+++ b/itk-module-init.cmake
@@ -32,14 +32,21 @@ endif()
 option(RTK_USE_CUDA "Use CUDA for RTK" ${RTK_USE_CUDA_DEFAULT})
 
 # Configure CUDA compilation options
+option(RTK_CUDA_VERSION "Specify the exact CUDA version that must be used for RTK")
 if(RTK_USE_CUDA)
   enable_language(CUDA)
   set(CMAKE_CUDA_RUNTIME_LIBRARY Static)
 
   include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  find_package(CUDAToolkit REQUIRED 8.0)
+  if(RTK_CUDA_VERSION)
+    find_package(CUDAToolkit EXACT ${RTK_CUDA_VERSION})
+  else()
+    find_package(CUDAToolkit REQUIRED 8.0)
+  endif()
 
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-declarations")
   set(RTK_CUDA_PROJECTIONS_SLAB_SIZE "16" CACHE STRING "Number of projections processed simultaneously in CUDA forward and back projections")
+elseif(RTK_CUDA_VERSION)
+  message(FATAL_ERROR "RTK_CUDA_VERSION is set but the CUDA toolkit has not been found.")
 endif()
diff --git a/setup.py b/setup.py
old mode 100644
new mode 100755
index f435017e6..3f9ff60ac
--- a/setup.py
+++ b/setup.py
@@ -12,10 +12,11 @@
 
 # Configure wheel name if CUDA is used
 wheel_name='itk-rtk'
-# Extract cuda version from the last folder name in CUDAToolkit_ROOT path.
+# Extract cuda version from the RTK_CUDA_VERSION cmake option
 for arg in sys.argv:
-  if "CUDAToolkit_ROOT" in str(arg):
-    wheel_name += ('-' + arg.rsplit('/', 1)[-1])
+  if "RTK_CUDA_VERSION" in str(arg):
+    cuda_version = arg.rsplit('RTK_CUDA_VERSION=', 1)[-1]
+    wheel_name += '-cuda' + cuda_version.replace('.', '')
 
 setup(
     name=wheel_name,