GU-Net: Low-Computation Diffuse Glioma Segmentation

Author: Arnav Dhar

A novel, simplified variant of BU-Net for segmenting diffuse gliomas (a prevalent adult brain tumor) in MRI scans, optimized specifically for low-computation neuroimaging.

Diffuse gliomas are currently identified and segmented manually by radiologists so that surgeons can use the result as a treatment reference — a slow process. Prior automated approaches use CNNs and U-Nets; one key piece of prior work, BU-Net, slightly alters the U-Net architecture. GU-Net simplifies BU-Net so it needs far less data and compute.

Results

Trained on a subset of the BraTS 2021 dataset — just 1,647 images from 549 brain MRIs — GU-Net achieves:

Metric	Score
Dice Similarity Coefficient (DSC)	71.58%
Intersection over Union (IoU)	61.29%

These are relatively high given the limited data and computation, and are comparable to BU-Net and U-Net despite using a significantly smaller portion of the data. The goal is to advance diagnosis in underprivileged areas and hospitals with limited funding, where less data and higher efficiency matter.

Architecture

og_gu_net.py contains the full model and training pipeline:

• Conv2DBlock — double convolution block (Conv → BN → ReLU, ×2).
• ES — multi-scale strip-convolution block with a residual skip.
• WC — factorized (width/height) convolution block used at the transition.
• BUNet — the GU-Net encoder/transition/decoder with average-pooling downsampling and transpose-convolution upsampling.
• DiceBCELoss, dice_coefficient, iou — combined Dice + BCE loss and evaluation metrics.
• MyDataset — loads preprocessed BraTS FLAIR/segmentation tensor slices.

Setup

pip install -r requirements.txt

Originally developed as a Google Colab notebook with Google Drive mounted. The training code expects the BraTS 2021 tensor slices under the path set in d_url and runs on CUDA.