diff --git a/Modules/Classification/CLActiveLearning/src/mitkActiveLearningInteractor.cpp b/Modules/Classification/CLActiveLearning/src/mitkActiveLearningInteractor.cpp index e85b5a3a87..c74ca8be65 100644 --- a/Modules/Classification/CLActiveLearning/src/mitkActiveLearningInteractor.cpp +++ b/Modules/Classification/CLActiveLearning/src/mitkActiveLearningInteractor.cpp @@ -1,549 +1,549 @@ /*=================================================================== The Medical Imaging Interaction Toolkit (MITK) Copyright (c) German Cancer Research Center, Division of Medical and Biological Informatics. All rights reserved. This software is distributed WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See LICENSE.txt or http://www.mitk.org for details. ===================================================================*/ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include // Helper function to get an image from a data node. static mitk::Image::Pointer GetImage(mitk::DataNode::Pointer dataNode) { if (dataNode.IsNull()) mitkThrow(); mitk::Image::Pointer image = dynamic_cast(dataNode->GetData()); if (image.IsNull()) mitkThrow(); return image; } // Helper function to get a geometry of an image for a specific time step. static mitk::BaseGeometry::Pointer GetGeometry(mitk::Image* image, unsigned int timeStep) { if (image == nullptr) mitkThrow(); mitk::TimeGeometry::Pointer timeGeometry = image->GetTimeGeometry(); if (timeGeometry.IsNull()) mitkThrow(); auto geometry = timeGeometry->GetGeometryForTimeStep(timeStep); if (geometry.IsNull()) mitkThrow(); return geometry; } static double EuclideanDistance(const mitk::Point3D p1, const mitk::Point3D p2) { return std::sqrt( (p1[0] - p2[0])*(p1[0] - p2[0]) + (p1[1] - p2[1])*(p1[1] - p2[1]) + (p1[2] - p2[2])*(p1[2] - p2[2]) ); } static double EuclideanDistance2D(const mitk::Point2D p1, const mitk::Point2D p2) { return std::sqrt( (p1[0] - p2[0])*(p1[0] - p2[0]) + (p1[1] - p2[1])*(p1[1] - p2[1]) ); } static double ScalarProduct(const mitk::Point3D p1, const mitk::Point3D p2) { return p1[0]*p2[0] + p1[1]*p2[1] + p1[2]*p2[2]; } static double ScalarProduct2D(const mitk::Point2D p1, const mitk::Point2D p2) { return p1[0]*p2[0] + p1[1]*p2[1]; } static std::vector> InterpolateIndices2D(itk::Index<2> startIndex, itk::Index<2> endIndex, const mitk::PlaneGeometry* geometry, unsigned int size) { if (geometry == nullptr) mitkThrow(); std::vector> resultIndices; mitk::Point2D startPoint; mitk::Point2D endPoint; startPoint[0] = startIndex[0]; startPoint[1] = startIndex[1]; endPoint[0] = endIndex[0]; endPoint[1] = endIndex[1]; geometry->IndexToWorld(startPoint, startPoint); geometry->IndexToWorld(endPoint, endPoint); // Distance between end points double dist = EuclideanDistance2D(startPoint, endPoint); // Define region between startIndex and endIndex, padded by (size - 1) int regionBounds[4]; for (int i=0; i<2; ++i) { regionBounds[2*i] = std::min(startIndex[i], endIndex[i]) - (size - 1); regionBounds[2*i+1] = std::max(startIndex[i], endIndex[i]) + (size - 1); } // We only have a threshold given in pixels (size), but image can be spaced in different units. // To get the corresponding distances, transform unit vectors and get their lengths. // The minimum spacing will be what we compare to. double spacings[2]; spacings[0] = geometry->GetExtentInMM(0) / geometry->GetExtent(0); spacings[1] = geometry->GetExtentInMM(1) / geometry->GetExtent(1); double minSpacing = (spacings[0] > spacings[1]) ? spacings[0] : spacings[1]; MITK_INFO << "Interpolating between " << startIndex << " and " << endIndex; MITK_INFO << "Distance: " << dist; MITK_INFO << "Minimum extent: " << minSpacing; double t = 0; double d = 0; for (int x = regionBounds[0]; x<=regionBounds[1]; ++x) { for (int y = regionBounds[2]; y<=regionBounds[3]; ++y) { mitk::Point2D p; mitk::Point2D index; index[0] = x; index[1] = y; geometry->IndexToWorld(index, p); // if there is not distance between start and end, just get distance to start if (dist < mitk::eps) { d = EuclideanDistance2D(startPoint, p); } else { t = -1./(dist*dist) * ScalarProduct2D(startPoint - p, endPoint - startPoint); if (t > 0. && t < 1.) { d = std::sqrt( ScalarProduct2D(startPoint - p, startPoint - p) + 2. * t * ScalarProduct2D(endPoint - startPoint, startPoint - p) + t * t * dist * dist ); } else if (t <= 0.) { d = EuclideanDistance2D(startPoint, p); } else { d = EuclideanDistance2D(endPoint, p); } } if (d <= minSpacing * static_cast(size) / 2.) { resultIndices.push_back(itk::Index<2>({index[0], index[1]})); } } } return resultIndices; } static std::vector> InterpolateIndices(itk::Index<3> startIndex, itk::Index<3> endIndex, const mitk::BaseGeometry* geometry, unsigned int size) { if (geometry == nullptr) mitkThrow(); std::vector> resultIndices; mitk::Point3D startPoint; mitk::Point3D endPoint; geometry->IndexToWorld(startIndex, startPoint); geometry->IndexToWorld(endIndex, endPoint); // Distance between end points double dist = EuclideanDistance(startPoint, endPoint); // Define region between startIndex and endIndex, padded by (size - 1) int regionBounds[6]; for (int i=0; i<3; ++i) { regionBounds[2*i] = std::min(startIndex[i], endIndex[i]) - (size - 1); regionBounds[2*i+1] = std::max(startIndex[i], endIndex[i]) + (size - 1); } // We only have a threshold given in pixels (size), but image can be spaced in different units. // To get the corresponding distances, transform unit vectors and get their lengths. // The minimum spacing will be what we compare to. // Could potentially use the spacing to normalize. double spacingInIndexSystem[3]; double minSpacing = -1.; for (int i=0; i<3; ++i) { itk::Index<3> origin; origin.Fill(0); itk::Index<3> index; index.Fill(0); index[i] = 1; mitk::Point3D p_origin; mitk::Point3D p_index; geometry->IndexToWorld(origin, p_origin); geometry->IndexToWorld(index, p_index); double spacing = EuclideanDistance(p_origin, p_index); if ( (minSpacing > 0. && spacing < minSpacing) || minSpacing < 0. ) { minSpacing = spacing; } spacingInIndexSystem[i] = spacing; } // Iterate over all indices in the given region and get distance to the line between startPoint and endPoint. // If distance is smaller than size, add to resultIndices. // // Let (x1,y1,z1) = startPoint, (x2,y2,z2) = endPoint, (x0,y0,z0) = p a point. // // Line is defined by: // [x1 + (x2-x1) * t] // v = [y1 + (y2-y1) * t] // [z1 + (z2-z1) * t] // // Then (with * dot product): // t(p) = - (startPoint - p) * (endPoint - startPoint) / |endPoint - startPoint|^2 // // And (with x cross product): // d(p) = |(p - startPoint) x (p - endPoint)| / |endPoint - startPoint| double t = 0; double d = 0; for (int x = regionBounds[0]; x<=regionBounds[1]; ++x) { for (int y = regionBounds[2]; y<=regionBounds[3]; ++y) { for (int z = regionBounds[4]; z<=regionBounds[5]; ++z) { mitk::Point3D p; itk::Index<3> index = {x,y,z}; geometry->IndexToWorld(index, p); // if there is not distance between start and end, just get distance to start if (dist < mitk::eps) { d = EuclideanDistance(startPoint, p); } else { t = -1./(dist*dist) * ScalarProduct(startPoint - p, endPoint - startPoint); if (t > 0. && t < 1.) { d = std::sqrt( ScalarProduct(startPoint - p, startPoint - p) + 2. * t * ScalarProduct(endPoint - startPoint, startPoint - p) + t * t * dist * dist ); } else if (t <= 0.) { d = EuclideanDistance(startPoint, p); } else { d = EuclideanDistance(endPoint, p); } } if (d <= minSpacing) { resultIndices.push_back(index); } } } } return resultIndices; // ======= OLD INTERPOLATION ======== // std::vector> resultIndices; // mitk::Point3D startPoint; // mitk::Point3D endPoint; // geometry->IndexToWorld(startIndex, startPoint); // geometry->IndexToWorld(endIndex, endPoint); // itk::Index<3> indexDelta; // int indexDeltaInc[3]; // for (int i=0; i<3; i++) // { // indexDelta[i] = endIndex[i] - startIndex[i]; // indexDeltaInc[i] = (indexDelta[i] > 0) ? 1 : (indexDelta[i] < 0) ? -1 : 0; // } // int argm[3] = {0, 1, 2}; // if (abs(indexDelta[1]) > abs(indexDelta[0])) // { // argm[0] = 1; // argm[1] = 0; // } // if (abs(indexDelta[2]) > abs(indexDelta[argm[1]])) // { // argm[2] = argm[1]; // argm[1] = 2; // } // if (abs(indexDelta[2]) > abs(indexDelta[argm[0]])) // { // argm[1] = argm[0]; // argm[0] = 2; // } // double slopes[2]; // slopes[0] = (endPoint[argm[1]] - startPoint[argm[1]]) / (endPoint[argm[0]] - startPoint[argm[0]]); // slopes[1] = (endPoint[argm[2]] - startPoint[argm[2]]) / sqrt((endPoint[argm[1]] - startPoint[argm[1]]) * (endPoint[argm[1]] - startPoint[argm[1]]) + (endPoint[argm[0]] - startPoint[argm[0]]) * (endPoint[argm[0]] - startPoint[argm[0]])); // itk::Index<3> currentIndex = startIndex; // mitk::Point3D currentPoint = startPoint; // while (currentIndex != endIndex) // { // currentIndex[argm[0]] += indexDeltaInc[argm[0]]; // geometry->IndexToWorld(currentIndex, currentPoint); // currentPoint[argm[1]] = startPoint[argm[1]] + slopes[0] * (currentPoint[argm[0]] - startPoint[argm[0]]); // currentPoint[argm[2]] = startPoint[argm[2]] + slopes[1] * sqrt((currentPoint[argm[1]] - startPoint[argm[1]]) * (currentPoint[argm[1]] - startPoint[argm[1]]) + (currentPoint[argm[0]] - startPoint[argm[0]]) * (currentPoint[argm[0]] - startPoint[argm[0]])); // geometry->WorldToIndex(currentPoint, currentIndex); // resultIndices.push_back(currentIndex); // } // return resultIndices; } mitk::ActiveLearningInteractor::ActiveLearningInteractor() : m_WorkingSlice(nullptr), m_WorkingPlane(nullptr), - m_Size(5), + m_Size(1), m_PaintingPixelValue(0) { } mitk::ActiveLearningInteractor::~ActiveLearningInteractor() { } void mitk::ActiveLearningInteractor::ConnectActionsAndFunctions() { CONNECT_FUNCTION("paint", Paint) CONNECT_FUNCTION("paint_interpolate", PaintInterpolate) CONNECT_FUNCTION("set_workingslice", SetWorkingSlice) CONNECT_FUNCTION("writeback_workingslice", WriteBackWorkingSlice) } void mitk::ActiveLearningInteractor::DataNodeChanged() { this->ResetToStartState(); } void mitk::ActiveLearningInteractor::SetWorkingSlice(mitk::StateMachineAction* /*action*/, mitk::InteractionEvent *event) { try { auto renderer = event->GetSender(); auto image = GetImage(this->GetDataNode()); auto timeStep = renderer->GetTimeStep(); auto geometry = GetGeometry(image, timeStep); auto planeGeometry = renderer->GetCurrentWorldPlaneGeometry(); // Check current plane geometry if (m_WorkingPlane.IsNotNull()) { bool isSameSlice (true); isSameSlice &= mitk::MatrixEqualElementWise(planeGeometry->GetIndexToWorldTransform()->GetMatrix(), m_WorkingPlane->GetIndexToWorldTransform()->GetMatrix()); isSameSlice &= mitk::Equal(planeGeometry->GetIndexToWorldTransform()->GetOffset(), m_WorkingPlane->GetIndexToWorldTransform()->GetOffset()); if (isSameSlice) return; } m_WorkingPlane = renderer->GetCurrentWorldPlaneGeometry()->Clone(); // Extract corresponding slice vtkSmartPointer reslice = vtkSmartPointer::New(); reslice->SetOverwriteMode(false); reslice->Modified(); auto extract = mitk::ExtractSliceFilter::New(reslice); extract->SetInput(image); extract->SetTimeStep(timeStep); extract->SetWorldGeometry(m_WorkingPlane); extract->SetResliceTransformByGeometry(geometry); extract->SetVtkOutputRequest(false); extract->Modified(); extract->Update(); m_WorkingSlice = extract->GetOutput(); } catch (itk::ExceptionObject& e) { mitkThrow() << "Could not set working slice, because:"; mitkThrow() << e.GetDescription(); } } void mitk::ActiveLearningInteractor::WriteBackWorkingSlice(mitk::StateMachineAction* /*action*/, mitk::InteractionEvent *event) { try { auto renderer = event->GetSender(); auto image = GetImage(this->GetDataNode()); auto timeStep = renderer->GetTimeStep(); auto geometry = GetGeometry(image, timeStep); // Write back vtkSmartPointer reslice = vtkSmartPointer::New(); reslice->SetInputSlice(m_WorkingSlice->GetVtkImageData()); reslice->SetOverwriteMode(true); reslice->Modified(); mitk::ExtractSliceFilter::Pointer insert = mitk::ExtractSliceFilter::New(reslice); insert->SetInput(image); insert->SetTimeStep(timeStep); insert->SetWorldGeometry(m_WorkingPlane); insert->SetResliceTransformByGeometry(geometry); insert->Modified(); insert->Update(); image->Modified(); this->GetDataNode()->Modified(); mitk::RenderingManager::GetInstance()->RequestUpdateAll(); } catch (itk::ExceptionObject& e) { mitkThrow() << "Could not set working slice, because:"; mitkThrow() << e.GetDescription(); } } void mitk::ActiveLearningInteractor::Paint(mitk::StateMachineAction* /*action*/, mitk::InteractionEvent* event) { if (m_PaintingPixelValue == -1) mitkThrow() << "Cannot paint negative values"; try { auto renderer = event->GetSender(); auto image = GetImage(this->GetDataNode()); auto timeStep = renderer->GetTimeStep(); auto geometry = GetGeometry(image, timeStep); auto positionEvent = dynamic_cast(event); auto position = positionEvent->GetPositionInWorld(); if (!geometry->IsInside(position)) return; // Okay, we're safe. Convert the mouse position to the index of the pixel // we're pointing at. itk::Index<3> index; itk::Index<3> oldIndex; geometry->WorldToIndex(position, index); geometry->WorldToIndex(m_LastPosition, oldIndex); // We don't need to paint over and over again while moving the mouse // pointer inside the same pixel. That's especially relevant when operating // on zoomed images. if (index != oldIndex) { // Convert index to slice geometry m_WorkingSlice->GetGeometry()->WorldToIndex(position, index); itk::Index<2> indexInPlane2D; indexInPlane2D[0] = index[0]; indexInPlane2D[1] = index[1]; // Get indices auto indices = InterpolateIndices2D(indexInPlane2D, indexInPlane2D, m_WorkingPlane, m_Size); // Fill indices mitk::ImagePixelWriteAccessor writeAccessor(m_WorkingSlice, m_WorkingSlice->GetSliceData(0)); for (auto i : indices) { writeAccessor.SetPixelByIndexSafe(i, m_PaintingPixelValue); } m_LastPosition = position; m_Used = true; } } catch (itk::ExceptionObject& e) { mitkThrow() << "Could not paint, because:"; mitkThrow() << e.GetDescription(); } } void mitk::ActiveLearningInteractor::PaintInterpolate(mitk::StateMachineAction* /*action*/, mitk::InteractionEvent* event) { if (m_PaintingPixelValue == -1) mitkThrow() << "Cannot paint negative values"; try { auto renderer = event->GetSender(); auto image = GetImage(this->GetDataNode()); auto timeStep = renderer->GetTimeStep(); auto geometry = GetGeometry(image, timeStep); auto positionEvent = dynamic_cast(event); auto position = positionEvent->GetPositionInWorld(); if (!geometry->IsInside(position)) return; // Okay, we're safe. Convert the mouse position to the index of the pixel // we're pointing at. itk::Index<3> index; itk::Index<3> oldIndex; geometry->WorldToIndex(position, index); geometry->WorldToIndex(m_LastPosition, oldIndex); // We don't need to paint over and over again while moving the mouse // pointer inside the same pixel. That's especially relevant when operating // on zoomed images. if (index != oldIndex) { // Convert index to slice geometry m_WorkingSlice->GetGeometry()->WorldToIndex(position, index); m_WorkingSlice->GetGeometry()->WorldToIndex(m_LastPosition, oldIndex); itk::Index<2> indexInPlane2D; itk::Index<2> oldIndexInPlane2D; indexInPlane2D[0] = index[0]; indexInPlane2D[1] = index[1]; oldIndexInPlane2D[0] = oldIndex[0]; oldIndexInPlane2D[1] = oldIndex[1]; // Get indices auto indices = InterpolateIndices2D(oldIndexInPlane2D, indexInPlane2D, m_WorkingPlane, m_Size); // Fill indices mitk::ImagePixelWriteAccessor writeAccessor(m_WorkingSlice, m_WorkingSlice->GetSliceData(0)); for (auto i : indices) { writeAccessor.SetPixelByIndexSafe(i, m_PaintingPixelValue); } m_LastPosition = position; m_Used = true; } } catch (itk::ExceptionObject& e) { mitkThrow() << "Could not paint with interpolation, because:"; mitkThrow() << e.GetDescription(); } } diff --git a/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.cpp b/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.cpp index f76d0724d3..eeb37c2c86 100644 --- a/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.cpp +++ b/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.cpp @@ -1,1229 +1,1279 @@ /*=================================================================== The Medical Imaging Interaction Toolkit (MITK) Copyright (c) German Cancer Research Center, Division of Medical and Biological Informatics. All rights reserved. This software is distributed WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See LICENSE.txt or http://www.mitk.org for details. ===================================================================*/ // Blueberry #include #include // Qt #include #include #include #include // Qmitk #include "QmitkActiveLearning.h" // MITK #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include +#include // ITK/VTK #include #include #include #include #include #include #include #include #include #include typedef ActiveLearning::FeaturePixelType FeaturePixelType; typedef ActiveLearning::AnnotationPixelType AnnotationPixelType; typedef ActiveLearning::LabelPixelType LabelPixelType; typedef ActiveLearning::FeatureMatrixType FeatureMatrixType; typedef ActiveLearning::LabelVectorType LabelVectorType; // Returns true if list has at least one entry and all entries are valid mitk::Images, otherwise false static bool SelectionAllImages(const QList& nodes) { if (nodes.empty()) { return false; } for (const auto& node : nodes) { if(!(node.IsNotNull() && dynamic_cast(node->GetData()) != nullptr)) return false; } return true; } // QColor to mitk::Color static mitk::Color QColorToMitkColor(const QColor& qcolor) { mitk::Color color; color.SetRed((float)qcolor.red() / 255); color.SetGreen((float)qcolor.green() / 255); color.SetBlue((float)qcolor.blue() / 255); return color; } // For debugging static void PrintAllLabels(mitk::LabelSetImage* image) { for (auto it=image->GetActiveLabelSet()->IteratorBegin(); it!=image->GetActiveLabelSet()->IteratorConstEnd(); ++it) { MITK_INFO << "Key: " << it->first << " - Name: " << it->second->GetName() << " - Value: " << it->second->GetValue() << " - Color: " << it->second->GetColor(); } } // Make values of labels a consistent range static void FillLabelValues(mitk::LabelSetImage* image) { int value(0); for (auto it=image->GetActiveLabelSet()->IteratorBegin(); it!=image->GetActiveLabelSet()->IteratorConstEnd(); ++it) { it->second->SetValue(value); value++; } image->GetActiveLabelSet()->SetActiveLabel(0); } // Fill image with zeros static void FillWithZeros(mitk::Image* image) { unsigned int size = image->GetPixelType().GetSize(); for (unsigned int i=0; iGetDimension(); i++) { size *= image->GetDimension(i); } for (unsigned int t=0; tGetTimeSteps(); t++) { mitk::ImageWriteAccessor accessor(image, image->GetVolumeData(0)); memset(accessor.GetData(), 0, size); } } template static Eigen::Matrix Transform(const std::vector images) { // Find size for output matrix [number of voxels, number of feature images] unsigned int size = images[0]->GetDimension(0); for (unsigned int i=1; i<3; ++i) { size *= images[0]->GetDimension(i); } Eigen::Matrix outputMatrix(size, images.size()); for (unsigned int i=0; i::Pointer imageItk; mitk::CastToItkImage(images[i], imageItk); outputMatrix.col(i) = Eigen::Matrix::Map(imageItk->GetBufferPointer(), size); } return outputMatrix; } template static mitk::Image::Pointer Transform(const Eigen::Matrix &inputMatrix, const mitk::Image::Pointer referenceImage) { typename itk::Image::Pointer imageItk; auto outputImage = mitk::Image::New(); outputImage->Initialize(mitk::MakeScalarPixelType(), *(referenceImage->GetTimeGeometry()->Clone())); mitk::CastToItkImage(outputImage, imageItk); auto it = itk::ImageRegionIterator>(imageItk, imageItk->GetLargestPossibleRegion()); int i = 0; while (!it.IsAtEnd()) { it.Set(inputMatrix(i, 0)); ++it; ++i; } mitk::GrabItkImageMemory(imageItk, outputImage); return outputImage; } template static std::vector Transform(const Eigen::Matrix &inputMatrix, const mitk::Image::Pointer referenceImage) { std::vector resultVector; for (int j=0; j::Pointer imageItk; auto outputImage = mitk::Image::New(); outputImage->Initialize(mitk::MakeScalarPixelType(), *(referenceImage->GetTimeGeometry()->Clone())); mitk::CastToItkImage(outputImage, imageItk); auto it = itk::ImageRegionIterator>(imageItk, imageItk->GetLargestPossibleRegion()); int i = 0; while (!it.IsAtEnd()) { it.Set(static_cast(inputMatrix(i, j))); ++it; ++i; } mitk::GrabItkImageMemory(imageItk, outputImage); resultVector.push_back(outputImage); } return resultVector; } template static void PrintMatrix(const Eigen::Matrix dataMatrix, int maxRows = 0) { if (maxRows == 0 || maxRows > dataMatrix.rows()) maxRows = dataMatrix.rows(); MITK_INFO << "---------------------"; for (int i=0; i static void PrintMatrix(const Eigen::Matrix dataMatrix, const Eigen::Matrix labelMatrix, int maxRows = 0) { if (labelMatrix.rows() < dataMatrix.rows()) return; if (maxRows == 0 || maxRows > dataMatrix.rows()) maxRows = dataMatrix.rows(); MITK_INFO << "---------------------"; for (int i=0; i static void GaussianSmoothing(const itk::Image* inputImage, const double sigma, mitk::Image::Pointer outputImage) { typedef itk::Image ImageType; typedef itk::Image FeatureImageType; auto filter = itk::DiscreteGaussianImageFilter::New(); filter->SetInput(inputImage); filter->SetVariance(sigma*sigma); filter->Update(); mitk::GrabItkImageMemory(filter->GetOutput(), outputImage); } template static void GaussianGradientMagnitude(const itk::Image* inputImage, const double sigma, mitk::Image::Pointer outputImage) { typedef itk::Image ImageType; typedef itk::Image FeatureImageType; auto filter = itk::GradientMagnitudeRecursiveGaussianImageFilter::New(); filter->SetInput(inputImage); filter->SetSigma(sigma); filter->Update(); mitk::GrabItkImageMemory(filter->GetOutput(), outputImage); } template static void LaplacianOfGaussian(const itk::Image* inputImage, const double sigma, mitk::Image::Pointer outputImage) { typedef itk::Image ImageType; typedef itk::Image FeatureImageType; auto filter = itk::LaplacianRecursiveGaussianImageFilter::New(); filter->SetInput(inputImage); filter->SetSigma(sigma); filter->Update(); mitk::GrabItkImageMemory(filter->GetOutput(), outputImage); } //template //static void StructureTensorEigenvalues(const itk::Image* inputImage, // float sigma, std::vector outputImages) //{ //// typedef itk::Image ImageType; //// auto filter = itk::StructureTensorEigenvalueImageFilter::New(); //// filter->SetInput(inputImage); //// filter->SetInnerScale(sigma); //// filter->SetOuterScale(sigma); //// filter->Update(); //// for (unsigned int i=0; iGetNumberOfOutputs(); i++) //// { //// mitk::GrabItkImageMemory(filter->GetOutput(i), outputImages[i]); //// } // typedef itk::Image ImageType; // typedef itk::Image> TensorImageType; // auto filter = itk::StructureTensorImageFilter::New(); // filter->SetInput(inputImage); // filter->SetNoiseScale(sigma); // filter->SetFeatureScale(sigma); // try // { // filter->Update(); // } // catch (...) // { // mitkThrow(); // } // std::vector eigenValueImages; // std::vector> eigenValueImageIterators; // for (unsigned int i=0; iSetRegions(inputImage->GetLargestPossibleRegion()); // image->Allocate(); // eigenValueImages.push_back(image); // itk::ImageRegionIterator it(image, image->GetLargestPossibleRegion()); // eigenValueImageIterators.push_back(it); // } // itk::ImageRegionConstIterator tensorImageIterator(filter->GetOutput(), filter->GetOutput()->GetLargestPossibleRegion()); // while (!tensorImageIterator.IsAtEnd()) // { // typename TensorImageType::PixelType::EigenValuesArrayType ev; // tensorImageIterator.Get().ComputeEigenValues(ev); // for (unsigned int i=0; i //static void HessianEigenvalues(const itk::Image* inputImage, // float sigma, std::vector outputImages) //{ // typedef itk::Image ImageType; // typedef itk::Image, imageDimension> TensorImageType; // auto filter = itk::HessianRecursiveGaussianImageFilter::New(); // ImageType::Pointer o1, o2, o3; // o1->Allocate(); // o2->Allocate(); // o3->Allocate(); // filter->SetInput(inputImage); // filter->SetSigma(sigma); // filter->Update(); // TensorImageType::Pointer tensorImage = filter->GetOutput(); // itk::ImageRegionIterator tensorIt(tensorImage, tensorImage->GetLargestPossibleRegion()); // itk::ImageRegionIterator o1It(o1, o1->GetLargestPossibleRegion()); // itk::ImageRegionIterator o2It(o2, o2->GetLargestPossibleRegion()); // itk::ImageRegionIterator o3It(o3, o3->GetLargestPossibleRegion()); // while (!tensorIt.IsAtEnd()) // { // itk::SymmetricSecondRankTensor::EigenValue // for (unsigned int i=0; iGetNumberOfOutputs(); i++) // { // mitk::GrabItkImageMemory(filter->GetOutput(i), outputImages[i]); // } //} /* ================================================================== * PUBLIC SLOTS * =============================================================== */ void ActiveLearning::Initialize() { // Get selected nodes and check again if these are all images m_Nodes = this->GetDataManagerSelection(); if (!SelectionAllImages(m_Nodes)) return; m_Controls.m_InitializePushButton->setDisabled(true); emit SignalSetProgressMaximum(m_Nodes.length() * 6 + 1); // 6 is number of features, shouldn't be hardcoded emit SignalResetProgress(); // Set names to the label (again) QString nameList = QString::fromStdString(m_Nodes[0]->GetName()); if (m_Nodes.length() >= 2) { for (int i=1; i"); nameList += QString::fromStdString(m_Nodes[i]->GetName()); } } m_Controls.m_InitializeLabel->setText(nameList); // ======================================= // PREDICTION NODE // ======================================= // m_PredictionNode = mitk::DataNode::New(); // m_PredictionNode->SetName("Predictions"); //// m_PredictionNode->SetColor(1., 1. ,1.); //// m_PredictionNode->SetBoolProperty("binary", false); // m_PredictionNode->SetProperty("opacity", mitk::FloatProperty::New(0.0f)); //// m_PredictionNode->SetBoolProperty("helper object", true); // this->GetDataStorage()->Add(m_PredictionNode, m_Nodes[0]); // ======================================= // SEGMENTATION IMAGE // ======================================= m_SegmentationImage = mitk::Image::New(); try { mitk::Image::Pointer referenceImage = dynamic_cast(m_Nodes[0]->GetData()); m_SegmentationImage->Initialize(mitk::MakeScalarPixelType(), *(referenceImage->GetTimeGeometry()->Clone())); } catch (mitk::Exception& e) { MITK_ERROR << "Exception caught: " << e.GetDescription(); QMessageBox::information(m_Parent, "Error", "Could not initialize segmentation image"); return; } FillWithZeros(m_SegmentationImage); m_SegmentationNode = mitk::DataNode::New(); m_SegmentationNode->SetData(m_SegmentationImage); m_SegmentationNode->SetName("Segmentation"); m_SegmentationNode->SetColor(1., 1., 1.); m_SegmentationNode->SetBoolProperty("binary", false); m_SegmentationNode->SetProperty("reslice interpolation", mitk::VtkResliceInterpolationProperty::New(VTK_RESLICE_NEAREST)); // m_PredictionNode->SetBoolProperty("helper object", true); this->GetDataStorage()->Add(m_SegmentationNode, m_Nodes[0]); // ======================================= // ANNOTATION IMAGE // ======================================= m_AnnotationImage = mitk::Image::New(); try { mitk::Image::Pointer referenceImage = dynamic_cast(m_Nodes[0]->GetData()); m_AnnotationImage->Initialize(mitk::MakeScalarPixelType(), *(referenceImage->GetTimeGeometry()->Clone())); } catch (mitk::Exception& e) { MITK_ERROR << "Exception caught: " << e.GetDescription(); QMessageBox::information(m_Parent, "Error", "Could not initialize annotation image"); return; } FillWithZeros(m_AnnotationImage); m_AnnotationNode = mitk::DataNode::New(); m_AnnotationNode->SetData(m_AnnotationImage); m_AnnotationNode->SetName("Labels"); m_AnnotationNode->SetColor(1., 1., 1.); m_AnnotationNode->SetBoolProperty("binary", false); m_AnnotationNode->SetProperty("reslice interpolation", mitk::VtkResliceInterpolationProperty::New(VTK_RESLICE_NEAREST)); m_AnnotationNode->SetProperty("opacity", mitk::FloatProperty::New(1.0f)); // m_AnnotationNode->SetBoolProperty("helper object", true); this->GetDataStorage()->Add(m_AnnotationNode, m_Nodes[0]); // Convert input images to FeaturePixelType for (auto node : m_Nodes) { mitk::Image::Pointer image = dynamic_cast(node->GetData()); auto itkImage = itk::Image::New(); mitk::CastToItkImage(image, itkImage); image = mitk::GrabItkImageMemory(itkImage); node->SetData(image); } emit SignalSetProgress(1); // Calculate features for (const auto node : m_Nodes) { mitk::Image::Pointer currentImage = dynamic_cast(node->GetData()); QFuture>> future; future = QtConcurrent::run(this, &ActiveLearning::CalculateFeatures, currentImage); auto futureWatcher = new QFutureWatcher>>(); futureWatcher->setFuture(future); connect(futureWatcher, SIGNAL(finished()), this, SLOT(OnInitializationFinished())); m_FeatureCalculationWatchers.push_back(futureWatcher); } // Interactor auto activeLearningLib = us::ModuleRegistry::GetModule("MitkCLActiveLearning"); m_Interactor = mitk::ActiveLearningInteractor::New(); m_Interactor->LoadStateMachine("Paint.xml", activeLearningLib); m_Interactor->SetEventConfig("PaintConfig.xml", activeLearningLib); m_Interactor->SetDataNode(m_AnnotationNode); // Automatically add first label OnAddLabelPushButtonClicked(); m_Active = true; } /* ================================================================== * PUBLIC * =============================================================== */ ActiveLearning::ActiveLearning() : m_Parent(nullptr), m_AnnotationImage(nullptr), m_AnnotationNode(nullptr), m_SegmentationImage(nullptr), m_SegmentationNode(nullptr), m_Active(false), m_NumberOfTrees(50), m_MaximumTreeDepth(10), m_SamplesPerTree(0.66), m_PredictionMatrix(nullptr) { } ActiveLearning::~ActiveLearning() { } void ActiveLearning::CreateQtPartControl( QWidget *parent ) { m_Controls.setupUi(parent); m_Parent = parent; // Label model m_LabelListModel = new QStandardItemModel(0, 3, this); m_Controls.m_LabelTableView->setModel(m_LabelListModel); m_Controls.m_LabelTableView->horizontalHeader()->setDefaultSectionSize(20); m_Controls.m_LabelTableView->verticalHeader()->setDefaultSectionSize(20); NotEditableDelegate* itemDelegate = new NotEditableDelegate(parent); m_Controls.m_LabelTableView->setItemDelegateForColumn(1, itemDelegate); // Connects connect(m_Controls.m_LabelTableView, SIGNAL(doubleClicked(QModelIndex)), this, SLOT(OnColorIconDoubleClicked(QModelIndex))); connect(m_Controls.m_LabelTableView->selectionModel(), SIGNAL(selectionChanged(QItemSelection, QItemSelection)), this, SLOT(OnLabelListSelectionChanged(QItemSelection, QItemSelection))); connect(m_LabelListModel, SIGNAL(dataChanged(QModelIndex, QModelIndex)), this, SLOT(OnLabelNameChanged(QModelIndex, QModelIndex))); connect(m_Controls.m_InitializePushButton, SIGNAL(clicked()), this, SLOT(Initialize())); connect(m_Controls.m_AddLabelPushButton, SIGNAL(clicked()), this, SLOT(OnAddLabelPushButtonClicked())); connect(m_Controls.m_RemoveLabelPushButton, SIGNAL(clicked()), this, SLOT(OnRemoveLabelPushButtonClicked())); connect(m_Controls.m_PaintToolButton, SIGNAL(clicked()), this, SLOT(OnPaintToolButtonClicked())); connect(m_Controls.m_EraseToolButton, SIGNAL(clicked()), this, SLOT(OnEraseToolButtonClicked())); connect(m_Controls.m_SaveSegmentationPushButton, SIGNAL(clicked()), this, SLOT(OnSaveSegmentationPushButtonClicked())); connect(m_Controls.m_SavePredictionsPushButton, SIGNAL(clicked()), this, SLOT(OnSavePredictionsPushButtonClicked())); connect(m_Controls.m_UpdatePredictionsPushButton, SIGNAL(clicked()), this, SLOT(OnUpdatePredictionsPushButtonClicked())); connect(this, SIGNAL(SignalIncrementProgress()), this, SLOT(OnSignalIncrementProgress())); connect(this, SIGNAL(SignalSetProgress(int)), this, SLOT(OnSignalSetProgress(int))); connect(this, SIGNAL(SignalResetProgress()), this, SLOT(OnSignalResetProgress())); connect(this, SIGNAL(SignalSetProgressMaximum(int)), this, SLOT(OnSignalSetProgressMaximum(int))); + connect(m_Controls.m_BrushSizeSlider, SIGNAL(valueChanged(int)), + this, SLOT(OnBrushSizeSliderValueChanged(int))); // Set start configuration m_Controls.m_LabelControlsFrame->setVisible(false); SetInitializeReady(false); } void ActiveLearning::ResetLabels() { for (int i=0; irowCount(); i++) { m_LabelListModel->item(i, 1)->setText(QString::number(i + 1)); } } std::vector > ActiveLearning::CalculateFeatures(const mitk::Image::Pointer inputImage) { std::vector> result; // TODO: Get features from preference page std::vector sigmas = {0.7, 1.6}; for (auto sigma : sigmas) { std::stringstream ss; auto gaussImage = mitk::Image::New(); AccessByItk_n(inputImage, GaussianSmoothing, (sigma, gaussImage)); gaussImage->SetClonedTimeGeometry(inputImage->GetTimeGeometry()); ss << "GaussianSmoothing (" << std::fixed << std::setprecision(2) << sigma << ")"; result.push_back(std::pair(gaussImage, ss.str())); ss.str(""); emit SignalIncrementProgress(); auto gradMagImage = mitk::Image::New(); AccessByItk_n(inputImage, GaussianGradientMagnitude, (sigma, gradMagImage)); gradMagImage->SetClonedTimeGeometry(inputImage->GetTimeGeometry()); ss << "GaussianGradientMagnitude (" << std::fixed << std::setprecision(2) << sigma << ")"; result.push_back(std::pair(gradMagImage, ss.str())); ss.str(""); emit SignalIncrementProgress(); auto logImage = mitk::Image::New(); AccessByItk_n(inputImage, LaplacianOfGaussian, (sigma, logImage)); logImage->SetClonedTimeGeometry(inputImage->GetTimeGeometry()); ss << "LaplacianOfGaussian (" << std::fixed << std::setprecision(2) << sigma << ")"; result.push_back(std::pair(logImage, ss.str())); ss.str(""); emit SignalIncrementProgress(); // auto structImage1 = mitk::Image::New(); // auto structImage2 = mitk::Image::New(); // auto structImage3 = mitk::Image::New(); // std::vector structImages = {structImage1, structImage2, structImage3}; // AccessByItk_n(inputImage, StructureTensorEigenvalues, (sigma, structImages)); // structImage1->SetClonedTimeGeometry(inputImage->GetTimeGeometry()); // structImage2->SetClonedTimeGeometry(inputImage->GetTimeGeometry()); // structImage3->SetClonedTimeGeometry(inputImage->GetTimeGeometry()); // ss << "StructureTensorEV1 (" << std::fixed << std::setprecision(2) << sigma << ")"; // result.push_back(std::pair(structImage1, ss.str())); // ss.str(""); // ss << "StructureTensorEV2 (" << std::fixed << std::setprecision(2) << sigma << ")"; // result.push_back(std::pair(structImage2, ss.str())); // ss.str(""); // if (inputImage->GetDimension() == 3) // { // ss << "StructureTensorEV3 (" << std::fixed << std::setprecision(2) << sigma << ")"; // result.push_back(std::pair(structImage3, ss.str())); // ss.str(""); // } // emit SignalIncrementProgress(); // auto hessianImage1 = mitk::Image::New(); // auto hessianImage2 = mitk::Image::New(); // auto hessianImage3 = mitk::Image::New(); // std::vector hessianImages = {hessianImage1, hessianImage2, hessianImage3}; // AccessByItk_n(inputImage, HessianEigenvalues, (sigma, hessianImages)); // ss << "HessianEigenvalue1 (" << std::fixed << std::setprecision(2) << sigma << ")"; // result.push_back(std::pair(hessianImage1, ss.str())); // ss.str(""); // ss << "HessianEigenvalue2 (" << std::fixed << std::setprecision(2) << sigma << ")"; // result.push_back(std::pair(hessianImage2, ss.str())); // ss.str(""); // if (inputImage->GetDimension() == 3) // { // ss << "HessianEigenvalue3 (" << std::fixed << std::setprecision(2) << sigma << ")"; // result.push_back(std::pair(hessianImage3, ss.str())); // ss.str(""); // } } return result; } std::pair> ActiveLearning::CalculatePrediction(const mitk::Image::Pointer annotationImage, const std::vector &featureImageVector, const mitk::Image::Pointer referenceImage, mitk::AbstractClassifier* classifier, std::shared_ptr predictionMatrix) { // Create prediction matrix if necessary if (predictionMatrix == nullptr) { FeatureMatrixType mat = Transform(featureImageVector); predictionMatrix = std::make_shared(mat); } emit SignalIncrementProgress(); // Get training data and train auto training = GetTrainingData(annotationImage, featureImageVector); classifier->Train(*training.second, *training.first); emit SignalIncrementProgress(); // Get result LabelVectorType segmentation = classifier->Predict(*predictionMatrix); emit SignalIncrementProgress(); FeatureMatrixType prediction = classifier->GetPointWiseProbabilities(); mitk::Image::Pointer segmentationImage = Transform(segmentation, referenceImage); std::vector predictionImages = Transform(prediction, referenceImage); emit SignalIncrementProgress(); std::pair> result = std::make_pair(segmentationImage, predictionImages); return result; } +mitk::Image::Pointer ActiveLearning::CalculateUncertainty(const std::vector &probabilityImageVector) +{ + typedef itk::Image ImageType; + std::vector itkImages; + std::vector> iterators; + + for (auto image : probabilityImageVector) + { + typename ImageType::Pointer itkImage; + mitk::CastToItkImage(image, itkImage); + itkImages.push_back(itkImage); + itk::ImageRegionConstIterator it(itkImage, itkImage->GetLargestPossibleRegion()); + iterators.push_back(it); + } + + auto uncertaintyImage = ImageType::New(); + uncertaintyImage->SetRegions(itkImages[0]->GetLargestPossibleRegion()); + uncertaintyImage->Allocate(); + itk::ImageRegionConstIterator uit(uncertaintyImage, uncertaintyImage->GetLargestPossibleRegion()); + + while (!uit.IsAtEnd()) + { + FeaturePixelType value = 0; + for (unsigned int i=0; i, std::shared_ptr> ActiveLearning::GetTrainingData(const mitk::Image::Pointer annotationImage, const std::vector &featureImageVector) { // Get indices and labels std::vector> indices; std::vector labels; itk::Image::Pointer annotationImageItk; mitk::CastToItkImage(annotationImage, annotationImageItk); itk::ImageRegionIteratorWithIndex> it(annotationImageItk, annotationImageItk->GetLargestPossibleRegion()); while (!it.IsAtEnd()) { if (it.Get() != 0) { indices.push_back(it.GetIndex()); labels.push_back(it.Get()); } ++it; } FeatureMatrixType trainingData(indices.size(), featureImageVector.size()); LabelVectorType trainingLabels = LabelVectorType::Map(labels.data(), labels.size()); int j = 0; for (mitk::Image::Pointer feature : featureImageVector) { int i = 0; mitk::ImagePixelReadAccessor access(feature, feature->GetVolumeData()); for (auto index : indices) { trainingData(i, j) = access.GetPixelByIndexSafe(index); i++; } j++; } auto trainingLabelsPtr = std::make_shared(trainingLabels); auto trainingDataPtr = std::make_shared(trainingData); std::pair, std::shared_ptr> result = std::make_pair(trainingLabelsPtr, trainingDataPtr); return result; } void ActiveLearning::UpdateLookupTables() { // Create new lookup table from list // Annotation type is int, but we only use a ushort lookup table auto lut = vtkSmartPointer::New(); int lowlim = std::numeric_limits::min(); int uplim = std::numeric_limits::max(); lut->SetNumberOfTableValues(uplim - lowlim + 1); lut->SetTableRange(lowlim, uplim); for (long i=0; i<(uplim-lowlim+1); ++i) { lut->SetTableValue(i, 0.0, 0.0, 0.0, 0.0); } for (int j=0; jrowCount(); ++j) { int value = m_LabelListModel->item(j, 1)->text().toInt(); const QColor color = m_LabelListModel->item(j, 0)->background().color(); lut->SetTableValue(value, color.redF(), color.greenF(), color.blueF(), 1.0); } auto lutMitk = mitk::LookupTable::New(); lutMitk->SetVtkLookupTable(lut); // Set to annotation image and segmentation image auto * lut_prop = dynamic_cast(m_AnnotationNode->GetProperty("LookupTable")); lut_prop->SetLookupTable(lutMitk); m_AnnotationNode->SetProperty("Image Rendering.Mode", mitk::RenderingModeProperty::New(mitk::RenderingModeProperty::LOOKUPTABLE_COLOR)); m_AnnotationNode->Modified(); lut_prop = dynamic_cast(m_SegmentationNode->GetProperty("LookupTable")); lut_prop->SetLookupTable(lutMitk); m_SegmentationNode->SetProperty("Image Rendering.Mode", mitk::RenderingModeProperty::New(mitk::RenderingModeProperty::LOOKUPTABLE_COLOR)); m_SegmentationNode->Modified(); } void ActiveLearning::UpdatePredictionNodes() { if (m_LabelListModel->rowCount() == 0) return; // Build lookup table vtkSmartPointer lut = vtkSmartPointer::New(); lut->SetTableRange (0, 1); lut->SetSaturationRange (0, 0); lut->SetHueRange (0, 0); lut->SetValueRange (0, 1); lut->SetAlphaRange (0, 1); lut->Build(); auto lutMitk = mitk::LookupTable::New(); lutMitk->SetVtkLookupTable(lut); for (int i=0; irowCount(); ++i) { auto property = mitk::NodePredicateProperty::New("segmentation_value", mitk::IntProperty::New(m_LabelListModel->item(i, 1)->text().toInt())); auto nodes = this->GetDataStorage()->GetDerivations(m_Nodes[0], property); if (nodes->Size() == 1) { auto node = nodes->GetElement(0); QString name = "Prediction "; name += m_LabelListModel->item(i, 2)->text(); node->SetName(name.toStdString()); auto * lut_prop = dynamic_cast(node->GetProperty("LookupTable")); lut_prop->SetLookupTable(lutMitk); node->SetProperty("Image Rendering.Mode", mitk::RenderingModeProperty::New(mitk::RenderingModeProperty::LOOKUPTABLE_COLOR)); node->SetColor(QColorToMitkColor(m_LabelListModel->item(i, 0)->background().color())); node->Modified(); } if (nodes->Size() > 1) mitkThrow(); } mitk::RenderingManager::GetInstance()->RequestUpdateAll(); } const std::string ActiveLearning::VIEW_ID = "org.mitk.views.activelearning"; /* ================================================================== * PROTECTED SLOTS * =============================================================== */ void ActiveLearning::OnSignalIncrementProgress() { m_Controls.m_ProgressBar->setValue(m_Controls.m_ProgressBar->value() + 1); } void ActiveLearning::OnSignalSetProgress(int value) { m_Controls.m_ProgressBar->setValue(value); } void ActiveLearning::OnSignalResetProgress() { m_Controls.m_ProgressBar->setValue(0); } void ActiveLearning::OnSignalSetProgressMaximum(int value) { m_Controls.m_ProgressBar->setMaximum(value); } void ActiveLearning::OnAddLabelPushButtonClicked() { QString labelName = QString("Label ") + QString::number(m_LabelListModel->rowCount() + 1); QColor labelColor = Qt::GlobalColor(m_LabelListModel->rowCount() % 12 + 7); // We only want Qt default colors 7 to 18 // Create icon QStandardItem* colorSquare = new QStandardItem; colorSquare->setBackground(labelColor); colorSquare->setEditable(false); QPixmap colorPixmap(20, 20); colorPixmap.fill(labelColor); colorSquare->setIcon(QIcon(colorPixmap)); // Key is the highest existing key + 1 int value = 1; if (m_LabelListModel->rowCount() >= 1) { value = m_LabelListModel->item(m_LabelListModel->rowCount() - 1, 1)->text().toInt() + 1; } QStandardItem* valueItem = new QStandardItem; valueItem->setText(QString::number(value)); // Create label item QStandardItem* label = new QStandardItem(labelName); // Make list and insert QList list; list.append(colorSquare); list.append(valueItem); list.append(label); m_LabelListModel->appendRow(list); m_Controls.m_LabelTableView->selectRow(m_LabelListModel->rowCount() - 1); // If this is the first label, we activate the paint button // We also have to set the data node color for this one, because for 1 values that color seems to define the rendered color if (m_LabelListModel->rowCount() == 1) { OnPaintToolButtonClicked(); } // Update colors UpdateLookupTables(); } void ActiveLearning::OnRemoveLabelPushButtonClicked() { // can't remove last label if (m_LabelListModel->rowCount() <= 1) return; QItemSelectionModel* selection = m_Controls.m_LabelTableView->selectionModel(); if (selection->hasSelection()) { unsigned int removeIndex = selection->selectedRows().first().row(); QString removeMessage = QString("Remove label '") + m_LabelListModel->item(removeIndex, 2)->text() + QString("'?"); QMessageBox::StandardButton removeReply; removeReply = QMessageBox::question(m_Parent, "Remove Label", removeMessage, QMessageBox::Yes | QMessageBox::No); if (removeReply == QMessageBox::Yes) { AnnotationPixelType removeValue = m_LabelListModel->item(removeIndex, 1)->text().toInt(); m_LabelListModel->removeRow(removeIndex); if (!m_Interactor->IsUsed()) { ResetLabels(); } else { itk::Image::Pointer imageItk; mitk::CastToItkImage(m_AnnotationImage, imageItk); auto it = itk::ImageRegionIterator>(imageItk, imageItk->GetLargestPossibleRegion()); while (!it.IsAtEnd()) { if (it.Get() == removeValue) it.Set(0); ++it; } mitk::GrabItkImageMemory(imageItk, m_AnnotationImage); UpdateLookupTables(); mitk::RenderingManager::GetInstance()->RequestUpdateAll(); } } } } void ActiveLearning::OnPaintToolButtonClicked() { m_Controls.m_PaintToolButton->setChecked(true); QItemSelectionModel* selection = m_Controls.m_LabelTableView->selectionModel(); int row(0); if (selection->hasSelection()) { row = selection->selectedRows().first().row(); } else { m_Controls.m_LabelTableView->selectRow(0); } m_Interactor->SetPaintingPixelValue(m_LabelListModel->item(row, 1)->text().toInt()); } void ActiveLearning::OnEraseToolButtonClicked() { m_Controls.m_EraseToolButton->setChecked(true); m_Interactor->SetPaintingPixelValue(0); } void ActiveLearning::OnSaveSegmentationPushButtonClicked() { auto newNode = mitk::DataNode::New(); newNode->SetName("Segmentation"); newNode->SetBoolProperty("binary", false); newNode->SetOpacity(1.0); newNode->SetVisibility(false); newNode->SetData(m_SegmentationImage->Clone()); this->GetDataStorage()->Add(newNode); } void ActiveLearning::OnSavePredictionsPushButtonClicked() { if (m_LabelListModel->rowCount() < 1) return; for (int i=0; irowCount(); ++i) { auto property = mitk::NodePredicateProperty::New("segmentation_value", mitk::IntProperty::New(m_LabelListModel->item(i, 1)->text().toInt())); auto nodes = this->GetDataStorage()->GetDerivations(m_Nodes[0], property); if (nodes->Size() == 1) { auto sourceNode = nodes->GetElement(0); mitk::Image::Pointer sourceImage = dynamic_cast(sourceNode->GetData()); auto newNode = mitk::DataNode::New(); QString name = "Prediction "; name += m_LabelListModel->item(i, 2)->text(); newNode->SetName(name.toStdString()); newNode->SetBoolProperty("binary", false); newNode->SetOpacity(1.0); newNode->SetVisibility(false); newNode->SetProperty("segmentation_value", mitk::IntProperty::New(m_LabelListModel->item(i, 1)->text().toInt())); newNode->SetData(sourceImage->Clone()); this->GetDataStorage()->Add(newNode); } if (nodes->Size() > 1) mitkThrow(); } } void ActiveLearning::OnColorIconDoubleClicked(const QModelIndex& index) { // Check if click is really from color icon if (index.column() != 0) { return; } else { // Color change dialog QColor setColor = QColorDialog::getColor(m_LabelListModel->itemFromIndex(index)->background().color(), m_Parent, "Select Label Color"); if (setColor.isValid()) { m_LabelListModel->itemFromIndex(index)->setBackground(setColor); QPixmap colorPixmap(20, 20); colorPixmap.fill(setColor); m_LabelListModel->itemFromIndex(index)->setIcon(QIcon(colorPixmap)); UpdateLookupTables(); UpdatePredictionNodes(); } } } void ActiveLearning::OnLabelListSelectionChanged(const QItemSelection& selected, const QItemSelection& /*deselected*/) { if (selected.empty()) return; if (m_Controls.m_EraseToolButton->isChecked()) return; // This assumes that only one item can be selected (single selection table view) try { int labelValue = m_LabelListModel->item(selected.indexes()[0].row(), 1)->text().toInt(); m_Interactor->SetPaintingPixelValue(labelValue); } catch (...) { m_Interactor->SetPaintingPixelValue(-1); } } void ActiveLearning::OnLabelNameChanged(const QModelIndex& topLeft, const QModelIndex& /*bottomRight*/) { UpdatePredictionNodes(); } void ActiveLearning::OnInitializationFinished() { // Check if all futures are finished for (auto watcher : m_FeatureCalculationWatchers) { if (watcher->isFinished() == false) {return;} } // Empty feature vector m_FeatureImageVector.clear(); // Insert features into feature vector and data storage for (unsigned int i=0; iresult(); for (unsigned int j=0; jSetData(result[j].first); node->SetName(result[j].second); node->SetBoolProperty("helper object", true); node->SetVisibility(false); this->GetDataStorage()->Add(node, m_Nodes[i]); } } // Show controls m_Controls.m_LabelControlsFrame->setVisible(true); m_Controls.m_InitializePushButton->setHidden(true); emit SignalResetProgress(); // Delete watchers for (auto watcher : m_FeatureCalculationWatchers) { delete watcher; } m_FeatureCalculationWatchers.clear(); } void ActiveLearning::OnUpdatePredictionsPushButtonClicked() { if (m_LabelListModel->rowCount() < 1) return; m_Controls.m_UpdatePredictionsPushButton->setDisabled(true); emit SignalSetProgressMaximum(4); // Clear old predictions for (int i=0; irowCount(); ++i) { auto property = mitk::NodePredicateProperty::New("segmentation_value", mitk::IntProperty::New(m_LabelListModel->item(i, 1)->text().toInt())); auto nodes = this->GetDataStorage()->GetDerivations(m_Nodes[0], property); if (nodes->Size() == 1) { this->GetDataStorage()->Remove(nodes); } if (nodes->Size() > 1) mitkThrow(); } // Classifier auto classifier = mitk::VigraRandomForestClassifier::New(); classifier->SetTreeCount(m_NumberOfTrees); classifier->SetMaximumTreeDepth(m_MaximumTreeDepth); classifier->SetSamplesPerTree(m_SamplesPerTree); mitk::Image::Pointer referenceImage = dynamic_cast(m_Nodes[0]->GetData()); QFuture>> future; future = QtConcurrent::run(this, &ActiveLearning::CalculatePrediction, m_AnnotationImage, m_FeatureImageVector, referenceImage, classifier, m_PredictionMatrix); m_PredictionCalculationWatcher = new QFutureWatcher>>(); m_PredictionCalculationWatcher->setFuture(future); connect(m_PredictionCalculationWatcher, SIGNAL(finished()), this, SLOT(OnPredictionCalculationFinished())); } void ActiveLearning::OnPredictionCalculationFinished() { auto result = m_PredictionCalculationWatcher->result(); m_SegmentationImage = result.first; m_SegmentationImage->Modified(); m_SegmentationNode->SetData(m_SegmentationImage); m_SegmentationNode->Modified(); for (unsigned int i=0; iitem(i, 2)->text(); node->SetName(name.toStdString()); node->SetBoolProperty("binary", false); node->SetVisibility(false); node->SetOpacity(0.3); node->SetColor(QColorToMitkColor(m_LabelListModel->item(i, 0)->background().color())); node->SetProperty("segmentation_value", mitk::IntProperty::New(m_LabelListModel->item(i, 1)->text().toInt())); node->SetData(result.second[i]); this->GetDataStorage()->Add(node, m_Nodes[0]); } UpdateLookupTables(); UpdatePredictionNodes(); emit SignalResetProgress(); m_Controls.m_UpdatePredictionsPushButton->setEnabled(true); } +void ActiveLearning::OnBrushSizeSliderValueChanged(int value) +{ + QString labelText = "Size "; + labelText += QString::number(value); + m_Controls.m_BrushSizeLabel->setText(labelText); + m_Interactor->SetSize(value); +} + /* ================================================================== * PROTECTED * =============================================================== */ void ActiveLearning::OnSelectionChanged(berry::IWorkbenchPart::Pointer /*source*/, const QList& nodes) { if (!SelectionAllImages(nodes)) { SetInitializeReady(false); return; } if (nodes.length() >= 2) { // First selection is the reference (could be any other) mitk::Image::Pointer referenceImage = dynamic_cast(nodes[0]->GetData()); mitk::BaseGeometry* referenceGeometry = referenceImage->GetTimeGeometry()->GetGeometryForTimeStep(0); // Adjust for multiple timesteps for (int i=1; i(nodes[i]->GetData()); mitk::BaseGeometry* currentGeometry = currentImage->GetTimeGeometry()->GetGeometryForTimeStep(0); // Adjust for multiple timesteps if (!mitk::Equal(*currentGeometry, *referenceGeometry, mitk::eps, true)) { SetInitializeReady(false); return; } } } // All nodes have the same geometry, allow init SetInitializeReady(true); } void ActiveLearning::SetFocus() { } /* ================================================================== * PRIVATE * =============================================================== */ void ActiveLearning::SetInitializeReady(bool ready) { if (ready) { // get selection, check again just to be sure auto nodes = this->GetDataManagerSelection(); if (!SelectionAllImages(nodes)) return; m_Controls.m_InitializePushButton->setEnabled(true); if (!m_Active) { QString nameList = QString::fromStdString(nodes[0]->GetName()); if (nodes.length() >= 2) { for (int i=1; i"); nameList += QString::fromStdString(nodes[i]->GetName()); } } m_Controls.m_InitializeLabel->setText(nameList); } } else { m_Controls.m_InitializePushButton->setDisabled(true); if (!m_Active) { m_Controls.m_InitializeLabel->setText("Selected images must have matching geometries"); } } } diff --git a/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.h b/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.h index 041c21f715..f9fbf6828b 100644 --- a/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.h +++ b/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearning.h @@ -1,186 +1,190 @@ /*=================================================================== The Medical Imaging Interaction Toolkit (MITK) Copyright (c) German Cancer Research Center, Division of Medical and Biological Informatics. All rights reserved. This software is distributed WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See LICENSE.txt or http://www.mitk.org for details. ===================================================================*/ #ifndef ActiveLearning_h #define ActiveLearning_h #include #include #include "ui_QmitkActiveLearningControls.h" // Qt #include #include #include #include // MITK #include #include #include #include /** \brief ActiveLearning \warning This class is not yet documented. Use "git blame" and ask the author to provide basic documentation. \sa QmitkAbstractView \ingroup ${plugin_target}_internal */ // Just a helper class class NotEditableDelegate : public QItemDelegate { Q_OBJECT public: explicit NotEditableDelegate(QObject* parent = nullptr) : QItemDelegate(parent) {} protected: QWidget* createEditor(QWidget*, const QStyleOptionViewItem&, const QModelIndex&) const {return Q_NULLPTR;} }; class ActiveLearning : public QmitkAbstractView { Q_OBJECT public slots: void Initialize(); public: typedef unsigned short AnnotationPixelType; typedef double FeaturePixelType; typedef int LabelPixelType; typedef Eigen::Matrix FeatureMatrixType; typedef Eigen::Matrix LabelVectorType; ActiveLearning(); ~ActiveLearning(); void CreateQtPartControl(QWidget *parent) override; std::vector> CalculateFeatures(const mitk::Image::Pointer inputImage); std::pair > CalculatePrediction(const mitk::Image::Pointer annotationImage, const std::vector &featureImageVector, const mitk::Image::Pointer referenceImage, mitk::AbstractClassifier *classifier, std::shared_ptr predictionMatrix); std::pair, std::shared_ptr > GetTrainingData(const mitk::Image::Pointer annotationImage, const std::vector &featureImageVector); + mitk::Image::Pointer CalculateUncertainty(const std::vector &probabilityImageVector); + void UpdateLookupTables(); void UpdatePredictionNodes(); static const std::string VIEW_ID; signals: void SignalIncrementProgress(); void SignalSetProgress(int value); void SignalResetProgress(); void SignalSetProgressMaximum(int value); protected slots: void OnSignalIncrementProgress(); void OnSignalSetProgress(int value); void OnSignalResetProgress(); void OnSignalSetProgressMaximum(int value); void OnAddLabelPushButtonClicked(); void OnRemoveLabelPushButtonClicked(); void OnPaintToolButtonClicked(); void OnEraseToolButtonClicked(); void OnSaveSegmentationPushButtonClicked(); void OnSavePredictionsPushButtonClicked(); void OnColorIconDoubleClicked(const QModelIndex& index); void OnLabelListSelectionChanged(const QItemSelection& selected, const QItemSelection& /*deselected*/); void OnLabelNameChanged(const QModelIndex& topLeft, const QModelIndex& /*bottomRight*/); + void OnBrushSizeSliderValueChanged(int value); + void OnUpdatePredictionsPushButtonClicked(); void OnPredictionCalculationFinished(); void OnInitializationFinished(); protected: void OnSelectionChanged(berry::IWorkbenchPart::Pointer /*source*/, const QList& nodes) override; void SetFocus() override; void ResetLabels(); void SetInitializeReady(bool ready); Ui::ActiveLearningControls m_Controls; QWidget* m_Parent; mitk::Image::Pointer m_AnnotationImage; mitk::DataNode::Pointer m_AnnotationNode; mitk::Image::Pointer m_SegmentationImage; mitk::DataNode::Pointer m_SegmentationNode; std::vector m_FeatureImageVector; std::vector>>*> m_FeatureCalculationWatchers; QFutureWatcher>>* m_PredictionCalculationWatcher; QStandardItemModel* m_LabelListModel; mitk::ActiveLearningInteractor::Pointer m_Interactor; QList m_Nodes; private: bool m_Active; int m_NumberOfTrees; int m_MaximumTreeDepth; float m_SamplesPerTree; std::shared_ptr m_PredictionMatrix; }; #endif // ActiveLearning_h diff --git a/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearningControls.ui b/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearningControls.ui index 23ba48a271..38308ae44a 100644 --- a/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearningControls.ui +++ b/Plugins/org.mitk.gui.qt.activelearning/src/internal/QmitkActiveLearningControls.ui @@ -1,491 +1,547 @@ ActiveLearningControls 0 0 352 581 0 0 QmitkTemplate Initialize with selection false false false false <font color="red">Selection must have matching dimensions and transforms</font> true QFrame::NoFrame QFrame::Raised 0 0 0 0 Qt::Vertical QSizePolicy::Fixed 20 10 QFrame::Sunken 1 Qt::Horizontal Qt::Vertical QSizePolicy::Fixed 20 10 QFrame::NoFrame QFrame::Raised 0 0 0 0 Add Label Remove Label Qt::Horizontal 40 20 QFrame::Sunken false QAbstractItemView::SingleSelection QAbstractItemView::SelectRows false true false QFrame::NoFrame QFrame::Raised 0 0 0 0 0 0 0 40 :/org.mitk.gui.qt.activelearning/resources/paint_icon_200.png:/org.mitk.gui.qt.activelearning/resources/paint_icon_200.png 25 25 true true 0 0 0 40 ... :/org.mitk.gui.qt.activelearning/resources/erase_icon_100.png:/org.mitk.gui.qt.activelearning/resources/erase_icon_100.png 25 25 true true + + + + QFrame::NoFrame + + + QFrame::Raised + + + + 0 + + + 0 + + + 0 + + + 0 + + + + + + 0 + 0 + + + + + 52 + 0 + + + + Size 1 + + + + + + + 1 + + + 20 + + + Qt::Horizontal + + + + + + Active Guidance 0 40 Update Qt::Vertical QSizePolicy::Fixed 20 10 Qt::Horizontal Qt::Vertical QSizePolicy::Fixed 20 10 Save to Data Manager QFrame::NoFrame QFrame::Raised 0 0 0 0 Segmentation Predictions QFrame::NoFrame QFrame::Raised 0 0 0 0 Qt::Vertical QSizePolicy::Fixed 20 10 Qt::Horizontal Qt::Vertical QSizePolicy::Fixed 20 10 0 false false Qt::Vertical 20 40