Hi everyone,
I wanted to address a problem I’ve been facing for the past week. I am currently working on a U-Net multiclass segmentation model, which works fine. The input is a (192, 192, 1) MRI image of a rat’s brain, and the output is a mask of size (192, 192, 3) with the three classes being: class 0 for background, class 1 for brain, and class 2 for lesion.
The metric I have been using is the Dice score. However, I wanted to add the Hausdorff distance, but so far, I haven’t been successful. Below is the code I’ve been using:
python
Copiar código
import numpy as np
from scipy.spatial.distance import cdist
def hausdorff_distance(set1, set2):
"""
Calculates the Hausdorff distance between two sets of points.
"""
# Calculate the distances between each point in set1 and all points in set2
dist1 = cdist(set1, set2, metric='euclidean')
dist2 = cdist(set2, set1, metric='euclidean')
# For each set, take the minimum distance to the other set
max_dist1 = np.max(np.min(dist1, axis=1)) # Maximum distance from each point in set1 to set2
max_dist2 = np.max(np.min(dist2, axis=1)) # Maximum distance from each point in set2 to set1
# The Hausdorff distance is the maximum of these two distances
return max(max_dist1, max_dist2)
def average_hausdorff_distance(y_true, y_pred, num_classes):
"""
Calculates the Average Hausdorff Distance for multiclass segmentations.
`y_true` and `y_pred` are the ground truth and predicted segmentation masks.
`num_classes` is the total number of classes in the segmentation.
"""
hausdorff_distances = []
# Iterate over each class
for class_id in range(num_classes):
# For each class, get the pixels of that class in the ground truth and predicted masks
true_class_points = np.argwhere(y_true[:, :, class_id] == 1) # Class 1
pred_class_points = np.argwhere(y_pred[:, :, class_id] == 1) # Class 1
if len(true_class_points) > 0 and len(pred_class_points) > 0:
hd = hausdorff_distance(true_class_points, pred_class_points)
hausdorff_distances.append(hd)
else:
hausdorff_distances.append(0) # If no points, append 0.
# Check if there are valid distances
if len(hausdorff_distances) == 0:
return 0.0 # If no distances can be calculated, return 0.
return np.mean(hausdorff_distances)
No matter the modifications or setups I have tried, the error remains the same: ValueError: None values not supported.
Any ideas? I couldn’t find any specific information online.
Thank you in advance