maryann-gitonga's picture
Update README.md
70c9874
metadata
tags:
  - semantic segmentation
  - brain tumor segmentation
library_name: tf
datasets:
  - https://www.kaggle.com/datasets/dschettler8845/brats-2021-task1
metrics:
  - Dice Coefficient
  - Tversky Loss
pipeline_tag: image-segmentation
license: cc-by-sa-4.0

3D Attention-based UNet for Multi-modal Brain Tumor Segmentation

Model Description

This model uses the UNet architecture, which employs a contracting path to down-sample image dimensions and an expanding path to up-sample while retaining spatial information through skip connections. 3D attention gates are introduced to generate 3D channel and spatial attention by utilizing 3D inter-channel and inter-spatial feature relationships. The input is a combined scan of 3 modalities (T1CE, T2 and T2-FLAIR) with the dimensions: 3 x L x W x no. of slices. The model attained a Dice Coefficient score of 0.9562 and a Tversky Loss of 0.0438

Dataset Description

The BRaTs (Brain Tumor Segmentation) 2021 Dataset, consisting of 1400 multi-parametric MRI (mpMRI) scans with expert neuro-radiologists' ground truth annotations, was used for this project. The dataset provides mpMRI scans in NIfTI format and includes native (T1), post-contrast T1-weighted (T1CE), T2-weighted (T2), and T2 Fluid Attenuated Inversion Recovery (T2-FLAIR) volumes, along with manually annotated GD-enhancing tumor, peritumoral edematous/invaded tissue, necrotic tumor core, and normal tissue. Checkout the dataset here.

Model Notebook

Find the notebook containing the model code on my Github.