This repository contains the Tensorflow implementation of our paper "CDANET: CHANNEL SPLIT DUAL ATTENTION BASED CNN FOR BRAIN TUMOR CLASSIFICATION IN MR IMAGES"
Python 3.7.13
Numpy 1.21.6
Tensorflow 2.8.0
Opencv-python 4.1.2
Pandas 1.3.5
h5py 3.1.0
Imgaug 0.2.9
-
Download the figshare dataset, extract the zip files and put them along with the cvind.mat in the kaggle/input folder.
-
To train the model run the following code: python train.py. Note that parameters and paths should be set beforehand
BraTS 2020 dataset contains two types of tumors: High-Grade Glioma(HGG) and Low-Grade Glioma(LGG). It consists of 293 HGG and 76 LGG 3D MR images. We used a stratified five fold cross validation scheme and trained the entire model in an end to end manner.
| Fold | Sensitivity | Specificity | F1-score | Accuracy |
|---|---|---|---|---|
| Fold1 | 96.87 | 98.10 | 97.40 | 97.78 |
| Fold2 | 96.59 | 97.40 | 96.89 | 96.61 |
| Fold3 | 95.64 | 96.54 | 96.00 | 96.32 |
| Fold4 | 97.03 | 96.00 | 96.49 | 97.29 |
| Fold5 | 94.58 | 94.35 | 94.33 | 95.02 |
| Average | 96.14 | 96.47 | 96.22 | 96.60 |
| Model | Attention Type | Sensitivity | Specificity | F1-Score | Accuracy |
|---|---|---|---|---|---|
| Densenet121 | Ours | 96.14 | 96.47 | 96.22 | 96.60 |
| Densenet121 | PAB | 95.72 | 95.58 | 95.43 | 95.95 |
| Densenet121 | CAB | 96.01 | 95.86 | 95.55 | 96.33 |
| Densenet121 | No Channel Split | 95.83 | 95.66 | 95.68 | 96.21 |
| Densenet121 | None | 95.34 | 95.28 | 95.09 | 95.88 |
| Author | Method | Accuracy(%) |
|---|---|---|
| Paul et al. | Custom CNN | 90.26 |
| Asfhar et al. | CapsNet | 86.56 |
| Swati et al. | VGG19 with BFT | 94.82 |
| Ghassemi et al. | GAN with custom CNN | 93.01 |
| Bodapati et al. | Two-Channel DNN | 95.23 |
| Bodapati et al. | MSENet | 95.37 |
| Abirami and Venkatesan | BCFA based GAN | 95.52 |
| Ours | CDANet | 96.60 |
| Fold | Sensitivity | Specificity | F1 Score | Accuracy |
|---|---|---|---|---|
| Fold1 | 88.31 | 80.00 | 83.94 | 82.69 |
| Fold2 | 100 | 100 | 100 | 100 |
| Fold3 | 100 | 100 | 100 | 100 |
| Fold4 | 100 | 100 | 100 | 100 |
| Fold5 | 100 | 100 | 100 | 100 |
| Avg. | 97.66 | 96.00 | 96.78 | 96.53 |
| Number of training folds used | Accuracy(in %) |
|---|---|
| 4 | 96.60 |
| 3 | 95.81 |
| 2 | 95.60 |
@INPROCEEDINGS{9897799,
author={Kumar Dutta, Tapas and Ranjan Nayak, Deepak},
booktitle={2022 IEEE International Conference on Image Processing (ICIP)},
title={CDANet: Channel Split Dual Attention Based CNN for Brain Tumor Classification In Mr Images},
year={2022},
volume={},
number={},
pages={4208-4212},
doi={10.1109/ICIP46576.2022.9897799}}