High-Throughput 3D Airway Segmentation — Reproducibility Package

Team 23 · 114-2 Parallel and Distributed Programming (NTU, Spring 2026) Paper: An Evolution From Heuristics to Neural Pipelines.

This package reproduces all results in the paper: four airway-segmentation pipelines compared on AIIB23 (120 chest-CT cases, one NVIDIA V100) under an identical I/O / timing harness. All four pipelines are one C++ binary (airway_seg) selected by a flag — the only differences are the post-processor and which (if any) neural model is loaded.

Pipeline	flag(s)	model	Paper result (Dice / s·case⁻¹)
Classical	--classical	none	0.6932 / 4.74
3D U-Net	--gate --gpu-softmax --no-stream-overlap	model/dynunet_retrained_fp16.onnx	0.9105 / 7.63
Hybrid	--hybrid --gpu-softmax --no-stream-overlap	model/dynunet_retrained_fp16.onnx	0.9074 / 7.67
2.5D U-Net	--model-2p5d --gate --gpu-softmax	model/airway_2p5d_unet.onnx	0.8634 / 4.01

The readable flags above are aliases baked into the binary. The old internal version tags (--v8/--v16/--v17/--unet25d/--v12-no-overlap) still work, so historical logs ([v16] gate FIRED, [v17] SAFETY ABORT) remain readable.

---

What's in this package

PDP_Team23_Reproducible/
├── README.md                  ← you are here (master guide)
├── env.sh                     ← cluster module loads (cuda + openmpi)
├── Makefile                   ← builds ./airway_seg from src/
├── src/                       ← SHARED C++/CUDA/MPI/ONNX engine (all 4 pipelines)
│
├── data/
│   ├── download_ct.sh         ← fetch AIIB23 CT + GT -> data/img, data/gt
│   ├── download_models.sh     ← fetch the 3D + 2.5D ONNX models -> model/
│   └── predecompress.sh       ← (optional) .nii.gz -> plain .nii for the fast load path
├── model/                     ← ONNX models (populated by download_models.sh) + README
├── third_party/
│   ├── libdeflate/            ← bundled (NIfTI gunzip)
│   └── download_onnxruntime.sh← fetch ONNX Runtime GPU 1.19.2 (~500 MB, not shipped)
│
├── pipelines/
│   ├── run_pipeline.sh        ← ONE dispatcher: run_pipeline.sh <pipeline> <cases|all>
│   └── README.md
│
├── training/                  ← reproduce the model WEIGHTS from scratch
│   ├── 3d_dynunet/            ← train.py + export_onnx.py (+ SLURM job)
│   └── 2p5d_unet/             ← prep_slices.py + train_unet25d.py (+ SLURM jobs)
│
├── analysis/                  ← CODE for every paper figure/table/stat (reads ../results/)
├── benchmark/                 ← SLURM harnesses for Tables II & III
├── results/                   ← canonical per-case CSVs behind the figures/tables
└── failed-opts/               ← rejected optimizations (Table IV) + negative analyses

---

Quick start

0. Prerequisites

• A Linux host with an NVIDIA GPU (paper used a V100, sm_70; CUDA ≥ 7.0).
• Toolchain: CUDA 12.x, OpenMPI 5.x (+UCX), mpicxx, nvcc, zlib. On TWCC Taiwania-2: source env.sh (module load cuda/12.8 openmpi/5.0.2_ucx1.14.1_cuda12.3).
• cuDNN 9 visible at link time (ships with a pip torch; see Makefile note).
• For training/analysis: Python 3.9+ with PyTorch + MONAI + pandas/matplotlib/scipy.

1. Get ONNX Runtime, the dataset, and the models

cd PDP_Team23_Reproducible

bash third_party/download_onnxruntime.sh      # ONNX Runtime GPU 1.19.2 (build + run neural)
pip install gdown
bash data/download_ct.sh                       # AIIB23 -> data/img/, data/gt/  (~50 GB)
bash data/download_models.sh                   # 3D + 2.5D ONNX -> model/        (~62 MB)

# (optional, matches the paper's 0.20 s load) pre-decompress to plain .nii
bash data/predecompress.sh data/img /work/$USER/decompressed_nii

Model Drive folder: https://drive.google.com/drive/folders/1PbvYLDMz3ETFT11TCaGkAl0Ly1MKjyd4

2. Build the engine

source env.sh                 # load modules (skip off-cluster if mpicxx/nvcc are on PATH)
make                          # -> ./airway_seg
# off-cluster you may need to point the linker at your cuDNN:
#   make CUDNN_RPATH=$(python -c "import nvidia.cudnn,os;print(os.path.dirname(nvidia.cudnn.__file__)+'/lib')")

3. Run any pipeline

bash pipelines/run_pipeline.sh classical 110       # Classical, one case
bash pipelines/run_pipeline.sh unet3d    110       # 3D U-Net
bash pipelines/run_pipeline.sh hybrid    168       # Hybrid (case 168 = the safety-abort case)
bash pipelines/run_pipeline.sh unet25d   all       # 2.5D U-Net, full 120-case benchmark

Each prints a per-stage timing block and Dice=...; masks land in results/<pipeline>/. See pipelines/README.md.

4. Regenerate the paper's figures/tables

cd analysis
python plot_accuracy_speed.py     # Dice vs speed (4 pipelines)
python plot_pervoxel.py           # per-voxel cost (38 vs 15 ns/voxel)
python plot_stage_breakdown.py    # per-stage wall-clock
python plot_mps_util.py           # GPU saturation
python stats_tests.py             # Wilcoxon + Cliff's delta
python stats_heldout_timing.py    # held-out val Dice + timing mean/std/95% CI

These read the cleaned per-case CSVs in results/ (our final runs), so they reproduce 4.74/7.63/7.67/4.01 (Dice 0.693/0.910/0.907/0.863) without a GPU. To re-measure on a GPU see benchmark/README.md (Tables II & III). Rejected optimizations (Table IV / "what did not help") are in failed-opts/README.md.

---

Reproducing the model weights (optional)

The neural pipelines load ready-to-run ONNX (download_models.sh), but both can be retrained:

• 3D DynUNet → training/3d_dynunet/README.md (train.py → export_onnx.py; held-out val Dice 0.8984).
• 2.5D U-Net → training/2p5d_unet/README.md (prep_slices.py → train_unet25d.py; held-out val Dice 0.841).

---

What you must supply vs. what is fetched

Item	In repo?	How to get it
C++/CUDA engine source + Makefile	✅	src/, Makefile
Result CSVs for figures/tables	✅	results/
libdeflate	✅	third_party/libdeflate/
3D + 2.5D inference ONNX (fp16)	❌ (~62 MB)	data/download_models.sh
ONNX Runtime GPU 1.19.2	❌ (~500 MB)	third_party/download_onnxruntime.sh
AIIB23 dataset	❌ (~50 GB)	https://codalab.lisn.upsaclay.fr/competitions/13238
Full .pth/.pt training checkpoints	❌	Google Drive (only to resume training)

---

Environment used for the paper

NVIDIA V100-SXM2 32 GB, driver 535.161.08 · CUDA 12.8 · OpenMPI 5.0.2 (UCX 1.14.1) · ONNX Runtime 1.19.2 with the TensorRT 10 EP in FP16 · TWCC Taiwania-2 (account mst114552). Inputs pre-decompressed to plain .nii; Dice evaluated against expert labels over all 120 AIIB23 cases.