Brain Tumor Multiclass Classification 🧠 - Computer Vision Model using a Convolutional Neuronal Network
This project focuses on creating a convolutional neural network designed to classify brain tumor MRI scans into three categories:
Gliomas, meningiomas and pituitary tumors.
The purpose of this model is to assist neuroradiologists in diagnosing these brain tumor types more efficiently, potentially speeding up the diagnostic process. However, it's important to emphasize that this model should complement human expertise rather than replace it entirely.
- Python 3.11
- Jupyter Notebook
pip install notebook - Required dependencies
pip install -r requirements.txt - Brain Tumor Image Dataset (1)
The original data set derived from 233 patients and was utilized for scientific publications and later converted to PNG images and published on Kaggle (1, 2).
Total number of MRI scans: 3064
Glioma images: 1426 (46.5%), Meningioma images: 708 (23.1%), Pituitary tumor images: 930 (30.4%)
The data set was split into a training set (80%, 2451 records), a validation set (10%, 306 records) and a test set (10%, 307 records).
For this task, a Convolutional Neural Network (CNN) known as ResNet50 was employed as the pre-trained base (3). It was prior compared against other CNN models like VGG16 and VGG19, showing a significantly superior performance. ResNet50 has 50 layers and makes use of a technique called residual learning. This involves using shortcuts/skip connections between layers to learn adjustments to the input data rather than attempting to directly learn the entire transformation. This approach makes the network very effective.
Randomly selected neurons are dropped out or ignored with a certain probability. This encourages generalization as the network does not rely too heavily on individual neurons.
L1/Lasso regularization, adds a 'penalty' to the absolute values of the weights to the loss. In this way, the neuronal network shows an increasing sparsity by driving some weights to zero, improving the capability to generalize.
L2/Ridge regularization adds a 'penalty' to the squared values of the weights to the loss, thereby encouraging smaller weight values and reducing the impact of individual features.
During training, the model's performance on the validation set is monitored. The training is stopped and the weights are restored when the performance on the validation set diminishes. This prevents the model from learning noise in the training data (4).
Several performance valuations were conducted to gain a detailed perspective on the predictions made with the test data:
❗ 
Accuracy is a good performance parameter when all classes are equally important and there are no significant class imbalances. In this project, classes were not well-balanced.
Precision is good when minimizing false positive predictions is important.
Recall is a good choice when minimizing false negative predictions is crucial.
F1-Score is suitable for situations where both false positives and false negatives are important to consider and when precision and recall are balanced (4).
Matrix of actual vs predicted classes for every instance. Each cell contains a count of how many instances were classified into a particular combination of actual and predicted classes. The diagonal represents the classes which were correctly classified.
(1) DENIZ KAVI. Brain Tumor Image Dataset, Retrieved 2/2024 from Kaggle
(2) Cheng, Jun, et al. "Enhanced Performance of Brain Tumor Classification via Tumor Region Augmentation and Partition." PloS one 10.10 (2015). doi: 10.1371/journal.pone.0140381.
(3) Kaiming H. et al. "Deep Residual Learning for Image Recognition". 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). URL.
(4) Géron, A. (2019). Hands-on machine learning with Scikit-Learn, Keras and TensorFlow: concepts, tools, and techniques to build intelligent systems (2nd ed.). O’Reilly.