This project was designed by Katlego Reuben KGOSI, and Mishi MAKADE, as our end year Bachelors degree in Data Science Capstone Project. The project classifies brain tumor MRI images, and segments the images. This project was motivated by seeing that EfficientNetB7 is one the best common pre-trained transfer learning models, but it hasnt been implemented in brain tumor classification. so we decided to compare it with state of the art most common best pre-trained transfer learning model VGG16. In this project we compare VGG16 and EfficientNetB7 in relation to accuracy and complexity and time of training and prediction. And implement the Best model in a web application.
EfficientNetB7,VGG16,classification,segmentation,transfer learning,pre-trained model
AUTHORS
- Katlego KGOSI, Sol Plaatje University, South Africa
- Mishi MAKADE, Sol Plaatje University, South Africa
| Table of Content |
|---|
| Specifications and requirements |
| Content of project with results |
| Acknowledgements |
| Web Application |
Machine Specifications
| Table of Content | Description |
|---|---|
Operating system |
Windows 10 & Windows 11 64-bit operating system |
RAM |
8GB & 16GB |
Graphics |
Intel(R) Core(TM) i5-8265U CPU @ 1.60GHz 1.80 GHz |
Editors |
Jupyter Notebbok, Spyder, Google Colab |
Requirements
- Python
- Tensorflow
- Keras
- matplotlib
- Streamlit
1. Data Collection & Pre-Processing
The data was downloaded from kaggle, Read the README file to get the link for the dataset.
A Glimpse of the data
Pre-Processing of the data
The images where transformed in the following manner
| Transformation | OutPut |
|---|---|
| Original Image | |
| horizontal flipping | |
| vertical flipping | |
| zooming at 0.2 | |
| rotation at 20 degrees | |
| featurewise_std_normalization | |
| shear-range of 0.2 | |
| brightening at the range 0.2-1.5 | |
| width shift range of 0.1 | |
| height shift range of 0.1 |
2. Model Development and Fine-tuning
The following parameters namely: Batch size, Epochs, Dropout and Optimizer, of the model where fined, and the visualisation of the results are below
The following values and parameters where chosen:
-
Optimizers - We will use only 5 optimizers to see which one works best, namely;
1. SGD,2. RMSprop,3. Adam,4. Adagrad,5. Adadelta.the best optimizer will be used in the New model -
Number of epochs - We will also use only 4 epochs to choose the best performer namely;
1, 2, 5, 10. -
Batch size - we will also use these batch sizes to choose the optimum batch size, namely;
8, 16. -
Dropout - we will optimise the dropout of the model, values
0.5, 0.6, 0.7, 0.8, 0.9
3. Final Model Implementation and Results
This is the structure of the EfficientNetB7 model we implemented with optimum parameters, similarly with the VGG16 model.
4. Conclusions
Our aim of our project was to compare VGG16 and EfficientNetB7, in terms of accuracy, complexity, and training time & prediction time(how long it takes to predict),
And our study shows that EfficientNetB7 takes longer to predict than VGG16, and also its training Time takes longer, and the best model in terms of accuracy is EfficientNetB7 with 98.19% over VGG16 which is 92.30%, therefore for our web app we chose EfficientNetb7 even though it has a limitation of taking longer to predict other VGG16.
I appreciate and acknowledge the supervision of the project by Mr Kudakwashe MADZIMA, my Capstone-Project Lecture at Sol Plaatje University in 2021, and also my collaborater Miss Mishi MAKADE, as we designed and implemented the project together.
will add the section of the Web App soon...
Thank you.