wenmin-wu/cv-black-slice-replacement

Black 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

1. Load each slice from the series
2. Check if image.mean() == 0 (all-black)
3. If black, randomly sample another slice from the same series
4. Retry up to N times to avoid infinite loops on fully corrupt volumes
5. Stack non-black slices into the output volume

Key Decisions

• Detection threshold: mean == 0 catches fully black; use mean < epsilon for 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

References

• [TF]: 3D & 2D Model for Brain Tumor Classification