Source code for monte_carlo_analysis.metrics.AUCPRClassWiseMetric

"""
**Author** : Robin Camarasa

**Institution** : Erasmus Medical Center

**Position** : PhD student

**Contact** : r.camarasa@erasmusmc.nl

**Date** : 2020-12-09

**Project** : monte_carlo_analysis

**Implement class AUCPRClassWiseMetric**

"""
from monte_carlo_analysis.metrics import Metric
import numpy as np
from sklearn.metrics import average_precision_score
from numba import jit


[docs]class AUCPRClassWiseMetric(Metric):
    """Implement AUCPRClassWiseMetric. This is a class-wise version of the :mod:`monte_carlo_analysis.metrics.AUCPRMetric`. In this scenario, the only misclassified voxels considered are those where the predicted class or the ground truth class is the class under study, all other voxels are considered correctly classified.
    """
    def __init__(self):
        super(AUCPRClassWiseMetric, self)

    def __call__(
            self, prediction: np.array,
            uncertainty_map: np.array,
            ground_truth: np.array
            ) -> list:
        """
        Compute the AUC-PR metric class-wise
        """
        # Get the number of classes
        nb_classes = prediction.shape[1]

        # Get the class prediction
        class_prediction = np.argmax(prediction.mean(axis=0), axis=0)
        class_ground_truth = np.argmax(ground_truth, axis=0)

        # Compute misclassification map
        results = []
        for i in range(nb_classes):
            class_ground_truth_ = (class_ground_truth == i)
            class_prediction_ = (class_prediction == i)
            misclassification_map = (class_ground_truth_ != class_prediction_)
            results.append(
                    average_precision_score(
                        misclassification_map.ravel(), uncertainty_map[i].ravel()
                        )
                    )
        return results