Python 3D numpy数组脑肿瘤分割后处理：移除阈值以下连通组件

Erosion-dilation is tricky for 3D medical imaging tasks like this—you risk shrinking valid small tumors or reconnecting noise after dilation. Instead, explicit connected component analysis gives you precise control over which regions to keep based on pixel count. Here's how to implement this for your 208x208x155 numpy array with classes 0/1/2/4:

Step-by-Step Implementation #

First, we'll use scikit-image's 3D-compatible tools to analyze each tumor class separately (since we don't want to mix components across different tumor types):

1. Import Required Libraries #

import numpy as np
from skimage.measure import label, regionprops

2. Define a Reusable Processing Function #

This function will iterate over each non-background class, identify connected components, and filter out any smaller than your 1000-pixel threshold:

def filter_small_3d_components(segmentation_arr, min_pixel_count=1000):
    # Create a copy to avoid modifying the original array
    filtered_arr = np.copy(segmentation_arr)
    
    # Iterate over each tumor class (skip background class 0)
    for class_label in [1, 2, 4]:
        # Create a binary mask for the current class
        class_mask = (filtered_arr == class_label)
        
        # Label all connected components in the 3D mask
        # connectivity=3 means 26-neighborhood in 3D (full 3D connectivity)
        labeled_components = label(class_mask, connectivity=3)
        
        # Get properties of each connected component
        for region in regionprops(labeled_components):
            # Remove components smaller than the threshold
            if region.area < min_pixel_count:
                filtered_arr[labeled_components == region.label] = 0
    
    return filtered_arr

3. Apply the Filter to Your Segmentation Output #

# Assuming your CNN output is stored in a variable called `cnn_segmentation`
filtered_segmentation = filter_small_3d_components(cnn_segmentation, min_pixel_count=1000)

Why This Works Better Than Erosion-Dilation #

• Precision: Unlike morphological operations, this targets only components that fall below your exact size threshold—no unintended shrinking or merging of valid regions.
• Class-Aware: It processes each tumor class independently, so you don't accidentally remove small regions from one class while keeping others.
• 3D-Centric: The connectivity=3 parameter uses full 3D neighborhood (26 adjacent voxels), which is critical for capturing true 3D tumor structures (erosion-dilation often uses isotropic kernels that might not match the anatomy).

Quick Notes #

• Make sure you're using a recent version of scikit-image (v0.17+ fully supports 3D connected component analysis).
• If your array is stored as a floating-point type, convert it to integer first with cnn_segmentation = cnn_segmentation.astype(np.int32) to avoid issues with labeling.
• Adjust connectivity if you need a different neighborhood (e.g., connectivity=1 for 6-neighborhood in 3D, but 26 is more common for medical imaging).

内容的提问来源于stack exchange，提问作者Himanshu