Skip to content
プロジェクト

leohsuofnthu/Pytorch-IterativeFCN

GitHubで見る

leohsuofnthu/Pytorch-IterativeFCN has 68 GitHub stars and 24 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-20 18:51 UTC. Last push 2026-09-28.

この数値の測り方

計測したスター増加

24時間
0
7日間
0
15日間
履歴を収集中
1か月
履歴を収集中

期間ごとの純増数です。期間は重複しており、日別の値ではありません。未計測の値は空欄です。スター数は注目度の指標であり、品質の保証ではありません。

概要

Pytorch implementation of the paper Iterative fully convolutional neural networks for automatic vertebra segmentation accepted in MIDL2018.

機械学習

Iterative FCN — Automatic Vertebra Segmentation (PyTorch)

PyTorch implementation of Iterative fully convolutional neural networks for automatic vertebra segmentation (MIDL 2018). A later journal version with anatomical identification is arXiv:1804.04383; this repo follows the MIDL-style iterative segmentation + completeness head (no anatomical label regression branch).

Can you retrain / run inference?

Yes. Pipeline:

  1. Install deps
  2. Preprocess CSI / SpineWeb Dataset 2 CT + labels → isotropic cropped volumes
  3. python train.py ... (or use weights/IterativeFCN_pretrained.pth)
  4. python instance_segmentation.py ... then optional python eval.py ...

You need a GPU with ~12 GB VRAM for 128³ patches (paper: Titan X, batch size 1).


Quick start

git clone https://github.com/leohsuofnthu/Pytorch-IterativeFCN.git
cd Pytorch-IterativeFCN

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
pip install -r requirements.txt

# Optional sanity checks (loads pretrained weights if present)
python -m pytest test/test_e2e.py -v

# Preprocess CSI dataset (see Dataset below)
python -m data.preprocessing --dataset ./CSI_dataset

# Train
python train.py --dataset ./crop_isotropic_dataset --iterations 80000 --log_interval 1000

# Instance segmentation on test images
python instance_segmentation.py \
  --test_dir ./crop_isotropic_dataset/test/img \
  --weights ./weights/IterativeFCN_pretrained.pth \
  --output_dir ./pred

# Dice vs labels
python eval.py --label_dir ./crop_isotropic_dataset/test/seg --pred_dir ./pred

Paper findings (audited against arXiv:1804.04383 PDF)

Topic Paper This repo
Input Image patch + instance memory Same (torch.cat of img + ins)
Patch size 128³, isotropic 1 mm³ Same
Segmentation head Binary current-instance mask + sigmoid (not softmax) Same — intentional (GitHub issue #6)
Completeness Sigmoid score + BCE Same
Seg loss λ·FP_soft + FN_soft with surface weights ω=γ·exp(-d²/σ²)+1, γ=8,σ=6 Same weights; λ now anneals 0.1→1 per paper
Optimizer Adam, lr 0.001, momentum 0.99, batch 1 Adam betas=(0.99,0.999), lr schedule 1e-3→1e-4→1e-5 (documented difference vs fixed 1e-3)
Inference Sliding window, vmin=1000, CoM refinement until converge Same + max_steps safety against endless loops
Anatomical ID loss In journal version Not implemented (MIDL / this repo)

Sigmoid vs softmax: the network labels “current vertebra vs not” for one instance at a time; sigmoid is correct. Softmax multi-class would be a different design.

Empty patches: training samples an empty-mask example every empty_interval=5 iterations (dataset) so the model can predict “nothing to segment”.


Dataset

CSI 2014 spine segmentation challenge (SpineWeb Dataset 2):
http://spineweb.digitalimaginggroup.ca/spineweb/index.php?n=Main.Datasets#Dataset_2.3A_Spine_and_Vertebrae_Segmentation

Some users report the download is flaky or offline (issues #3/#4). You need CT volumes + vertebra label masks in MetaImage (.mhd/.raw) form for preprocessing.

Place raw files under e.g. ./CSI_dataset/, then:

python -m data.preprocessing \
  --dataset ./CSI_dataset \
  --output_isotropic ./isotropic_dataset \
  --output_crop ./crop_isotropic_dataset \
  --split_ratio 0.8

Preprocessing: resample to 1 mm³ → distance weight maps → crop unlabeled vertebrae → crop_isotropic_dataset/{train,test}/{img,seg,weight}.


Training

python train.py \
  --dataset ./crop_isotropic_dataset \
  --weight ./weights \
  --checkpoints ./checkpoints \
  --iterations 80000 \
  --log_interval 1000 \
  --eval_iters 20 \
  --lr 1e-3 \
  --num_channels 64
Flag Notes
--num_channels 64 Required to match IterativeFCN_pretrained.pth (old bug used 4)
--resume Restore checkpoints/latest_checkpoints.pth
--device cuda / cpu Auto-detected if omitted

Outputs: weights/IterativeFCN_best_train.pth, weights/IterativeFCN_best_valid.pth, checkpoints under ./checkpoints/.


Inference & evaluation

python instance_segmentation.py \
  --test_dir ./crop_isotropic_dataset/test/img \
  --weights ./weights/IterativeFCN_pretrained.pth \
  --output_dir ./pred \
  --patch_size 128 --min_vol 1000 --max_alter 20
python eval.py --label_dir ./crop_isotropic_dataset/test/seg --pred_dir ./pred

Model

model

Constant-width 3D U-Net-style FCN + completeness classifier. Released checkpoint shares encoder/decoder block weights (see iterativeFCN.py).

Segmentation examples

normal

empty

result

This repo (partial train) Paper
~0.918 Dice ~0.958 Dice

Tests

python -m pytest test/test_e2e.py -v

Covers forward shapes, sigmoid binary head, pretrained weight load, FP/FN loss, patch clamping.


Tips

Issue Fix
OOM Keep --batch-size 1; reduce other GPU load
Output size is too small Patches must be 128³ (divisible by 16); do not feed 64³
Endless instance loop Fixed with caudal z stop + --max_steps; update code
Pretrained load mismatch Use num_channels=64
Dataset unavailable See SpineWeb / issue tracker; preprocessing needs local CSI files

Authors

  • HSU, CHIH-CHAO — ML Master’s student at Mila

References