Interactive Brain Tumor Segmentation (BraTS)

This repository contains the code and implementation details for accurately segmenting images of the human brain, specifically targeting tumor diagnosis and treatment planning. The proposed method utilizes deep neural networks with interactive segmentation techniques, aiming to improve accuracy while minimizing user effort and computational requirements.

Documents

• Thesis - fa
• Slides - en

Table of Contents

1. Abstract
2. Key Features
3. Installation
4. Simple Run
5. Experiments and Results
6. License
7. Acknowledgement

Abstract

Accurately segmenting 3D images of the human brain is crucial for medical applications such as tumor diagnosis and treatment planning. Deep neural networks have emerged as the most advanced automatic segmentation method, but they may require modifications for clinical use. Interactive segmentation allows for greater accuracy by incorporating user interactions, but current methods are not suitable for low-power systems and may require significant user effort. This study utilizes a novel method that performs a coarse segmentation on a low-resolution section of the target region, followed by a local refinement to restore lost resolution. To optimize efficiency, morphological analysis is used to modify only areas that need updating while preserving previously obtained results for other regions. Experimental results demonstrate that this method achieves more accurate results with less user interaction and requires less computing power and time compared to other methods. Additionally, this method shows acceptable generalization in tasks not encountered during the training phase.

Key Features

• Deep learning-based approach for accurate brain image segmentation
• Interactive segmentation techniques to incorporate user interactions
• Coarse segmentation followed by local refinement to restore resolution
• Efficient computation using selective morphological analysis
• Reduced user effort and improved accuracy compared to existing methods
• Generalization capabilities beyond the training phase

Installation

To set up the project, please follow these steps:

1. Clone this repository:

git clone https://github.com/ali-sedaghi/Interactive-BraTS.git

2. Install the required dependencies:

pip install -r requirements.txt

3. Download the dataset (BraTS2020) from source link and place it in the following directory:

data/datasets/BraTS/

4. Download the pretrained models (SegFormer, HRNet, ResNet) from source link and place it in the following directory:

data/weights/

5. To train the model, run the following command with:

python ./scripts/train.py ./models/brats/segformerB3.py\
  --exp-name=Test\
  --epochs=2\
  --batch-size=2\
  --ngpus=0\
  --cpu\
  --workers=2\
  --datapath=./data/datasets/BraTS\
  --channel=mix\
  --label=wt\
  --pretrained

6. To evaluate the model, run the following command with:

python ./scripts/evaluate.py FocalClick\
  --model_dir=./experiments/brats/segformerB3/000_FocalClick-SegFormerB3/checkpoints/\
  --checkpoint=epoch-19-val-loss-1.28.pth\
  --infer-size=96\
  --datasets=BraTS_Val\
  --datapath=./data/datasets/BraTS\
  --channel=flair\
  --label=wt\
  --cpu\
  --target-iou=0.95\
  --n-clicks=20\
  --thresh=0.5\
  --print-ious\
  --vis

Refer to the source code and Thesis for additional options and configurations.

Simple Run

You can use our ready to use notebooks on Kaggle and Google Colab.

• Kaggle train and evaluation notebook: [Link]
• Google Colab train and evaluation notebook: [Link]
• Kaggle evaluation notebook: [Link]

Important hyperparameters:

MODE = "FocalClick"  # CDNet, FocalClick
BACKBONE = "segformerB3"  # cdnet_res34, hrnet18s, hrnet32, segformerB3
CHANNEL = "mix"  # flair, t1, t1ce, t2, mix
LABEL = "wt"  # net, ed, et, wt, tc
EPOCHS = 20
BATCH_SIZE = 32
# Set --pretrained in below cell

Experiments and Results

The experimental results demonstrate that our proposed method achieves more accurate segmentation results with less user interaction and requires reduced computing power and time compared to other methods. The model also exhibits acceptable generalization capabilities beyond the training phase.

For detailed results, analysis, and trained models please refer to our Experiments Drive and Thesis.

NoC and NoF metrics in Experiment 4

Metric	NoC80%	NoC85%	NoC90%	NoF85%	NoF90%	TPC	Time
IoU	3.28	4.03	5.75	320	530	0.059	0:55:07
Dice	2.45	2.81	3.48	176	246	0.059	0:55:07

mIoU and mDice metrics in Experiment 4

Metric	Clicks=1	Clicks=2	Clicks=3	Clicks=5	Clicks=10	Clicks=20
mIoU	85.35 %	85.30 %	86.46 %	88.37 %	90.29 %	91.75 %
mDice	88.96 %	88.84 %	89.75 %	91.16 %	92.78 %	94.12 %

Output generated by the algorithm during a training phase step in Experiment 4

Output generated by the algorithm during an evaluation phase step in Experiment 4

Retrieving image details from test samples

License

This project is licensed under the Apache License. See the LICENSE file for details.

Acknowledgement

The core framework of this codebase follows:

https://github.com/XavierCHEN34/ClickSEG