image of train test accuracy over epochs

CNN Classification of Brain Tumor MRI Images

Nov 28, 2023

Summary
Investigates the use of Convolutional Neural Networks (CNN) for classifying brain tumors in MRI scans. Developed a CNN model that classifies tumors using grayscale 2-dimensional MRI images, achieving a notable accuracy of 90.84%. This achievement is particularly significant as it was accomplished without the use of transfer learning or 3-dimensional data, leading to a much leaner model. The study employed data augmentation techniques to enhance model robustness and accuracy. Despite minor imbalances in the data categories, the model proved effective across various tumor types. The work highlights the potential of simple, efficient CNN architectures in medical imaging, offering a viable tool for aiding diagnosis in resource-constrained environments.
PDF

In the fight against brain tumors in Canada, the integration of Convolutional Neural Networks (CNNs) into MRI analysis represents a major leap forward. Our research focuses on leveraging the power of CNNs to classify brain tumors with high accuracy using 2D grayscale MRI images, simplifying the process without compromising effectiveness.

Figure 1 shows various transformed images created by ImageDataGenerator to enhance model robustness against overfitting.

Our innovative model bypasses the complexities and high costs associated with 3D data acquisition, instead training on hundreds of 2D images to detect different tumor types with an impressive 90.84% accuracy. This not only demonstrates the model's capability to capture complex features but also its potential as a reliable diagnostic tool that remains accessible even in resource-limited settings.

Figure 2 illustrates our CNN model's architecture, highlighting the progression from input through convolutional layers to the output layer.

Through extensive research and application of data augmentation techniques, we've refined our CNN to perform with accuracy that rivals models based on transfer learning. This ensures our CNN model remains computationally efficient while still delivering precise classification outcomes.

Figure 3. Analysis of training and validation accuracy over time for Model 10, illustrating the achieved improvement in performance.

The table of metrics below provides a comprehensive view of the model's precision, recall, and f1-score across different tumor types, affirming its reliability and precision in classification tasks.

Table I. Classification Report
__________________________________________________________
|              | Precision | Recall | f1-score | support |
|--------------|-----------|--------|----------|---------|
| glioma       | 1.00      | 0.76   | 0.86     | 380     |
| meningioma   | 0.86      | 0.92   | 0.89     | 301     |
| No tumor     | 0.85      | 1.00   | 0.92     | 381     |
| Pituitary    | 0.96      | 0.96   | 0.96     | 336     |
|--------------------------------------------------------|
| accuracy     |           |        | 0.91     | 1398    |
| macro avg    | 0.92      | 0.91   | 0.91     | 1398    |
| Weighted avg | 0.92      | 0.91   | 0.91     | 1398    |
----------------------------------------------------------

In conclusion, the study highlights that a streamlined CNN model, without the use of complex transfer learning techniques or demanding 3D data, is capable of classifying MRI brain scans with high accuracy. These findings pave the way for the deployment of advanced yet cost-effective diagnostic tools in clinical settings, offering a beacon of hope for enhancing patient outcomes through technological innovation.