cv-black-slice-replacement
Detect all-black DICOM/MRI slices (mean==0) and replace them by randomly sampling a non-black slice from the same series
SKILL.mdskills/cv/black-slice-replacement/SKILL.md
GitHubBlack Slice Replacement
Overview
Medical imaging volumes sometimes contain entirely black slices due to scanner artifacts, corrupt files, or padding. Feeding black slices into a model wastes capacity and can degrade training. Detecting slices with mean == 0 and replacing them with a randomly sampled non-black slice from the same series maintains the expected input shape while providing meaningful pixel data.
Quick Start
import random
import numpy as np
def load_slice_with_fallback(path, all_paths, read_fn, max_retries=100):
"""Load a DICOM slice; if all-black, replace with random non-black slice."""
image = read_fn(path)
retries = 0
while image.mean() == 0 and retries < max_retries:
image = read_fn(random.choice(all_paths))
retries += 1
return image
def load_volume_safe(file_paths, read_fn):
"""Load volume with black-slice replacement."""
slices = []
for p in file_paths:
slices.append(load_slice_with_fallback(p, file_paths, read_fn))
return np.stack(slices)
Workflow
- Load each slice from the series
- Check if
image.mean() == 0(all-black) - If black, randomly sample another slice from the same series
- Retry up to N times to avoid infinite loops on fully corrupt volumes
- Stack non-black slices into the output volume
Key Decisions
- Detection threshold:
mean == 0catches fully black; usemean < epsilonfor near-black - Replacement strategy: random sampling is simple; nearest non-black slice preserves spatial context better
- Max retries: cap at 100 to handle volumes where most slices are black
- vs zero-pad: replacement provides real texture; zero-pad is cleaner but less informative