brainseg-utils 0.1.2

Lightweight research utilities for BraTS-style brain tumor segmentation

brainseg-utils

A lightweight Python library for BraTS-style brain tumor segmentation research. Provides utilities for loading NIfTI volumes, extracting 2D slices, computing segmentation metrics, and building PyTorch datasets — without requiring Docker or heavy dependencies.

---

Motivation

When working on brain tumor segmentation projects using the BraTS dataset, researchers repeatedly write the same boilerplate code for every experiment: loading .nii.gz files, normalizing intensities, extracting non-empty slices, avoiding data leakage from slice-level splitting, and computing region-wise Dice scores. This library consolidates those utilities into a clean, minimal API.

---

Installation

pip install brainseg-utils

---

Features

• NIfTI volume loading and intensity normalization (z-score, minmax)
• 2D slice extraction along any axis with empty slice filtering
• Patient-level train/val/test splitting to prevent data leakage
• Dice, IoU, and Hausdorff distance metrics
• BraTS region-wise Dice (WT, TC, ET)
• Matplotlib overlay visualization
• PyTorch Dataset class for drop-in DataLoader integration

---

Usage

from brainseg_utils import load_brats_subject, normalize, extract_slices, brats_dice

# load all modalities for a subject
subject = load_brats_subject("path/to/brats/", "BraTS20_Training_001")

# normalize
vol = normalize(subject["flair"], method="zscore")

# extract non-empty axial slices
slices = extract_slices(vol, subject["seg"], skip_empty=True)

# compute region-wise dice
scores = brats_dice(pred, subject["seg"])
print(scores)  # {'WT': 0.89, 'TC': 0.76, 'ET': 0.71}

Patient-level split

from brainseg_utils import patient_split

train_ids, val_ids, test_ids = patient_split(all_subject_ids, train=0.7, val=0.15, test=0.15)

PyTorch Dataset

from brainseg_utils.dataset import BraTSSliceDataset
from torch.utils.data import DataLoader

dataset = BraTSSliceDataset(root="path/to/brats", subject_ids=train_ids, modality="flair")
loader = DataLoader(dataset, batch_size=8, shuffle=True)

Visualization

from brainseg_utils import show_slice

show_slice(vol[:, :, 80], mask[:, :, 80])

---

Dependencies

• nibabel
• numpy
• matplotlib
• scipy (optional, for Hausdorff distance)
• torch (optional, for Dataset class)

---

Author

Dipendra Thapa GitHub: https://github.com/Dipendra367 Portfolio: https://dipendra367.com.np

---

License

MIT