wenmin-wu/cv-quantile-slice-sampling

Quantile Slice Sampling

Overview

3D medical imaging models require fixed-depth input, but CT/MRI scans vary widely in slice count (50-500+). Quantile slice sampling selects N evenly-spaced slices by computing quantile indices over the full stack. This preserves anatomical coverage from top to bottom regardless of scan length, unlike truncation (loses anatomy) or padding (wastes compute).

Quick Start

import numpy as np
from glob import glob

def sample_slices(scan_dir, n_slices=64):
    paths = sorted(glob(f"{scan_dir}/*"),
                   key=lambda x: int(x.split('/')[-1].split('.')[0]))
    n_scans = len(paths)
    if n_scans <= n_slices:
        indices = list(range(n_scans))
    else:
        indices = np.quantile(
            list(range(n_scans)),
            np.linspace(0., 1., n_slices)
        ).round().astype(int)
    return [paths[i] for i in indices]

# Sample 64 slices from a scan with 300+ slices
selected = sample_slices("/data/train_images/patient_001", n_slices=64)

Workflow

1. List and sort all slice files in the scan directory
2. Compute N quantile positions uniformly from 0 to 1
3. Map quantile positions to integer indices in the slice array
4. Load only the selected slices and stack into a 3D volume

Key Decisions

• n_slices: Match model architecture input depth (32, 64, 128 typical)
• Sort order: Sort by DICOM InstanceNumber or filename to preserve anatomical order
• vs stride sampling: Quantile handles edge cases better (always includes first and last)
• Short scans: If scan has fewer slices than target, use all slices and pad

References

• RSNA 2022 1st Place Solution - Train Stage1