Brain MRI Segmentation Server Guide

Files

• unet.py: completed PyTorch U-Net baseline
• train.py: CLI training script
• evaluate.py: CLI evaluation / visualization script
• ablation_registry.py: ablation experiment definitions
• run_experiments.py: batch runner for baseline, ablations, and combined improvements
• run_experiments.sh: shell wrapper around run_experiments.py
• summarize_experiments.py: build ablation tables and training-curve plots
• outputs/: default output directory

Recommended server environment

Install PyTorch with the CUDA command from the official PyTorch site first, then install the remaining packages:

pip install medpy scikit-image pillow tqdm matplotlib tensorboard

torchvision is not required by the current scripts.

Dataset path

The scripts expect the raw patient folders under a path like:

/path/to/BrainMRI/kaggle_3m

If your server path is different, pass it with --images or IMAGES_DIR=....

Full server run

From Project_1_0506/brain-seg/pytorch:

bash run_experiments.sh

The default runner now executes:

• baseline
• bn
• bilinear
• dropout
• bce_dice
• adamw
• cosine
• strong_aug
• improved

The shared default configuration is aligned with train.ipynb baseline:

• IMAGE_SIZE=256
• EPOCHS=100
• STEPS_PER_EPOCH=0 for full epochs
• BATCH_SIZE=16
• EVAL_BATCH_SIZE=16
• INIT_FEATURES=32
• VALIDATION_CASES=10
• VIS_IMAGES=200
• VIS_FREQ=10

All of them can be overridden with environment variables:

IMAGE_SIZE=256 BATCH_SIZE=8 EPOCHS=30 WORKERS=8 bash run_experiments.sh

You can also run only a subset of experiments:

EXPERIMENTS=baseline,bn,bilinear,improved bash run_experiments.sh

If you need to resume after partial completion:

SKIP_EXISTING=1 bash run_experiments.sh

Manual commands

Baseline:

python train.py \
  --images /path/to/BrainMRI/kaggle_3m \
  --output-root ./outputs \
  --experiment-name baseline \
  --epochs 100 \
  --batch-size 16 \
  --eval-batch-size 16 \
  --image-size 256 \
  --validation-cases 10 \
  --init-features 32 \
  --optimizer adam \
  --loss dice \
  --lr 3e-4 \
  --vis-images 200 \
  --vis-freq 10 \
  --workers 4

Single ablation example:

python train.py \
  --images /path/to/BrainMRI/kaggle_3m \
  --output-root ./outputs \
  --experiment-name bn \
  --epochs 100 \
  --batch-size 16 \
  --eval-batch-size 16 \
  --image-size 256 \
  --validation-cases 10 \
  --init-features 32 \
  --optimizer adam \
  --loss dice \
  --lr 3e-4 \
  --batch-norm \
  --vis-images 200 \
  --vis-freq 10 \
  --workers 4

Combined improved:

python train.py \
  --images /path/to/BrainMRI/kaggle_3m \
  --output-root ./outputs \
  --experiment-name improved \
  --epochs 100 \
  --batch-size 16 \
  --eval-batch-size 16 \
  --image-size 256 \
  --validation-cases 10 \
  --init-features 32 \
  --optimizer adamw \
  --loss bce_dice \
  --bce-weight 0.4 \
  --lr 1e-3 \
  --weight-decay 1e-4 \
  --scheduler cosine \
  --batch-norm \
  --bilinear \
  --dropout 0.1 \
  --aug-scale 0.10 \
  --aug-angle 20 \
  --vis-images 200 \
  --vis-freq 10 \
  --workers 4

Evaluation:

python evaluate.py \
  --images /path/to/BrainMRI/kaggle_3m \
  --experiment-dir ./outputs/baseline \
  --batch-size 16 \
  --workers 4

Outputs

For each experiment directory:

• config.json: training configuration
• history.csv: epoch metrics
• metrics.json: best validation summary
• weights/best.pt: best checkpoint
• evaluation/metrics.json: final evaluation summary
• evaluation/dsc_distribution.png: patient-level DSC plot
• evaluation/best_*.png, median_*.png, worst_*.png: overlay examples

After all experiments finish, the runner also creates:

• outputs/summary/ablation_results.csv
• outputs/summary/ablation_results.md
• outputs/summary/ablation_results.json
• outputs/summary/train_loss_curves.png
• outputs/summary/valid_loss_curves.png
• outputs/summary/valid_dsc_curves.png
• outputs/summary/best_dsc_bar.png

Suggested workflow for your report

1. Run bash run_experiments.sh.
2. Read the ablation table from outputs/summary/ablation_results.csv.
3. Insert the training curves from outputs/summary/*.png.
4. Use representative overlays from each experiment's evaluation/ directory when needed.

If GPU memory is limited

Reduce one or more of these:

• BATCH_SIZE
• IMAGE_SIZE
• INIT_FEATURES

Example:

IMAGE_SIZE=96 BATCH_SIZE=4 INIT_FEATURES=16 bash run_experiments.sh