AsymmetryNet: Asymmetry Attention Mechanism (AAM)
Segmentation-free feature extraction for radiologist-inspired left-right
asymmetry in head and neck CT, developed for HPV status prediction in
oropharyngeal squamous cell carcinoma (OPSCC).
What this code does
Given a cropped neck CT volume (NIfTI), this pipeline computes three
per-slice asymmetry channels across the full volume in a single pass:
- Ch1 — Airway lesion: seeded region growing from the airway lumen,
bounded by Sobel-detected soft-tissue edges, constrained to the side
of airway deviation from a spine-anchored midline.
- Ch2 — Nodal/soft-tissue asymmetry: mirrored left-right comparison
of fat-plane and soft-tissue density about the same midline.
- Ch3 — Necrosis: per-case, histogram-derived HU-windowed fluid
detection, restricted to the union of the Ch1 and Ch2 regions.
Midline is detected per case via spine-anchored bone-mass/compactness
scoring, not image center or a fixed landmark.
Outputs (per case)
- A QC plot (5 representative slices x 5 panels: CT+midline+spine,
Sobel-cleaned edge map, and the three asymmetry channels).
- A 4D heatmap NIfTI (
D x H x W x 3, channels = [airway_lesion, nodal,
necrosis]) for downstream radiomic feature extraction (e.g., MIRP)
restricted to these asymmetry-defined regions.
- A row of ~40 rich per-case features (areas, extents, laterality
ratios, 3D weighted centroids per channel) appended to a CSV.
Usage
Set BASE, CROP_DIR, and MASTER at the top of the script to your
own paths. Input volumes are expected as normalized (0-1) or raw-HU
NIfTI crops at {CROP_DIR}/{dataset}/{case_id}/crop224.nii.gz; the
master index CSV needs case_id, dataset, extracted (bool), and
optionally hpv_norm columns.
Parallelized via ProcessPoolExecutor and resumable: interrupting and
restarting skips cases already written to the output CSV.
Citation
If you use this code, please cite:
[full citation once accepted/published]
License
MIT
