wenmin-wu/cv-modality-adaptive-dicom-windowing

Overview

Mixed-modality medical datasets like RSNA Intracranial Aneurysm contain CT, CTA, MRA, and MRI series in the same training set. A single fixed windowing (or per-image min-max) collapses contrast on whichever modality wasn't tuned for. The fix: read Modality from the DICOM header, look it up in a {modality: (window_center, window_width)} table, and apply the matching window before scaling to uint8. The result is that vessel structure on CTA, flow signal on MRA, and parenchyma on MRI each land in a comparable normalized range — and one shared model can treat all of them as a "vessel image" rather than learning modality-specific intensity hacks.

Quick Start

import numpy as np
import pydicom

WINDOWS = {
    'CT':  (40,   80),     # parenchyma
    'CTA': (50,   350),    # vessels
    'MRA': (600,  1200),   # bright-blood angio
    'MRI': (40,   80),     # generic MR
}

def window(img, modality):
    wc, ww = WINDOWS.get(modality, (40, 80))
    lo, hi = wc - ww // 2, wc + ww // 2
    img = np.clip(img, lo, hi)
    return ((img - lo) / (hi - lo + 1e-7) * 255).astype(np.uint8)

ds = pydicom.dcmread(fp, force=True)
px = ds.pixel_array.astype(np.float32) * float(ds.RescaleSlope) + float(ds.RescaleIntercept)
img = window(px, getattr(ds, 'Modality', 'CT'))

Workflow

1. Inventory unique Modality strings in your training set (CT, CTA, MR, MRA)
2. Look up clinically-correct window center/width per modality (radiology atlases or kaggle host docs)
3. Always apply RescaleSlope * px + RescaleIntercept first — CT only, but harmless on MR
4. Window then scale to [0, 255] uint8 (or [0, 1] float)
5. Pass through the same downstream loader/augmentation as if all modalities were a single "image"
6. As a feature, also keep a one-hot modality embedding fed to the head — gives the model an explicit signal it cares about per-modality differences

Key Decisions

• Window first, normalize second: per-image min-max on raw HU is dominated by air/bone outliers and destroys soft-tissue contrast.
• Rescale slope/intercept before windowing: stored pixel values are not HU until rescaled.
• MR has no HU: clip to per-volume percentiles (1, 99) instead of fixed numbers if the host hasn't provided a window.
• Per-modality vs. multi-window stack: single window per modality is simpler; if you have GPU budget, stack 3 windows per modality for a richer input.
• Always carry the modality string forward: it's a strong feature for any classification head.
• Hardcoded windows beat data-driven: clinical windows are robust; learned ones overfit per-fold.

References

• RSNA-IAD | Ensemble | LB #1