inv_label_time_course.cpp

Go to the documentation of this file.

//=============================================================================================================
 
//=============================================================================================================
// INCLUDES
//=============================================================================================================
 
#include "inv_label_time_course.h"
 
//=============================================================================================================
// EIGEN INCLUDES
//=============================================================================================================
 
#include <Eigen/Core>
#include <Eigen/SVD>
 
//=============================================================================================================
// STD INCLUDES
//=============================================================================================================
 
#include <cmath>
#include <algorithm>
 
//=============================================================================================================
// QT INCLUDES
//=============================================================================================================
 
#include <QDebug>
#include <QMap>
 
//=============================================================================================================
// USED NAMESPACES
//=============================================================================================================
 
using namespace INVLIB;
using namespace FSLIB;
using namespace Eigen;
 
//=============================================================================================================
// DEFINE MEMBER METHODS
//=============================================================================================================
 
VectorXd InvLabelTimeCourse::computeSignFlip(const MatrixXd& stcData)
{
    const int nVerts = static_cast<int>(stcData.rows());
    if (nVerts == 0)
        return VectorXd();
 
    // Use SVD to find the dominant temporal pattern
    // Then flip each vertex to align with it
    JacobiSVD<MatrixXd> svd(stcData, ComputeThinU);
    VectorXd u1 = svd.matrixU().col(0);
 
    VectorXd signs(nVerts);
    for (int i = 0; i < nVerts; ++i)
        signs[i] = (u1[i] >= 0.0) ? 1.0 : -1.0;
 
    return signs;
}
 
//=============================================================================================================
 
MatrixXd InvLabelTimeCourse::extract(const InvSourceEstimate& stc,
                                     const QList<FsLabel>& labels,
                                     const QString& sMode,
                                     bool bAllowEmpty)
{
    if (stc.isEmpty() || labels.isEmpty()) {
        qWarning() << "[InvLabelTimeCourse::extract] Empty stc or labels.";
        return MatrixXd();
    }
 
    const int nTimes = static_cast<int>(stc.data.cols());
    const int nLabels = labels.size();
 
    // Resolve "auto" mode
    QString mode = sMode;
    if (mode == "auto")
        mode = "mean_flip";
 
    // Build a map from vertex index → row in stc.data
    QMap<int, int> vertexToRow;
    for (int i = 0; i < stc.vertices.size(); ++i)
        vertexToRow.insert(stc.vertices[i], i);
 
    // Count output labels (skip empty if !bAllowEmpty)
    QList<int> outputLabelIndices;
    for (int li = 0; li < nLabels; ++li) {
        int count = 0;
        for (int vi = 0; vi < labels[li].vertices.size(); ++vi) {
            if (vertexToRow.contains(labels[li].vertices[vi]))
                ++count;
        }
        if (count > 0 || bAllowEmpty)
            outputLabelIndices.append(li);
    }
 
    MatrixXd result = MatrixXd::Zero(outputLabelIndices.size(), nTimes);
 
    for (int oi = 0; oi < outputLabelIndices.size(); ++oi) {
        const FsLabel& label = labels[outputLabelIndices[oi]];
 
        // Find matching vertex rows
        QList<int> rows;
        for (int vi = 0; vi < label.vertices.size(); ++vi) {
            auto it = vertexToRow.find(label.vertices[vi]);
            if (it != vertexToRow.end())
                rows.append(it.value());
        }
 
        if (rows.isEmpty())
            continue;
 
        const int nVerts = rows.size();
 
        // Extract sub-matrix for this label
        MatrixXd labelData(nVerts, nTimes);
        for (int i = 0; i < nVerts; ++i)
            labelData.row(i) = stc.data.row(rows[i]);
 
        if (mode == "mean") {
            result.row(oi) = labelData.colwise().mean();
        } else if (mode == "mean_flip") {
            VectorXd signs = computeSignFlip(labelData);
            MatrixXd flipped = labelData;
            for (int i = 0; i < nVerts; ++i)
                flipped.row(i) *= signs[i];
            result.row(oi) = flipped.colwise().mean();
        } else if (mode == "pca_flip") {
            // Demean across time
            RowVectorXd meanTime = labelData.colwise().mean();
            MatrixXd centered = labelData.rowwise() - meanTime;
 
            JacobiSVD<MatrixXd> svd(centered, ComputeThinV);
            // First PC time course
            RowVectorXd pc1 = svd.matrixV().col(0).transpose();
 
            // Scale by singular value and sign
            double sv1 = svd.singularValues()[0];
            pc1 *= sv1 / std::sqrt(static_cast<double>(nVerts));
 
            // Flip sign to match the dominant sign of the data
            // Check correlation with mean
            double corr = (meanTime.array() * pc1.array()).sum();
            if (corr < 0.0)
                pc1 = -pc1;
 
            result.row(oi) = pc1;
        } else if (mode == "max") {
            // Maximum absolute value at each time point
            for (int t = 0; t < nTimes; ++t) {
                double maxVal = 0.0;
                double maxAbsVal = 0.0;
                for (int i = 0; i < nVerts; ++i) {
                    double absVal = std::abs(labelData(i, t));
                    if (absVal > maxAbsVal) {
                        maxAbsVal = absVal;
                        maxVal = labelData(i, t);
                    }
                }
                result(oi, t) = maxVal;
            }
        } else {
            qWarning() << "[InvLabelTimeCourse::extract] Unknown mode:" << mode
                       << "— using mean.";
            result.row(oi) = labelData.colwise().mean();
        }
    }
 
    return result;
}