The idea of the project was to reproduce and improve the result of a scientific publication on Brain tumor detection using CNN (Convolutional Neural Network) with MRI images.
The CNN will extract features maps from each image through an operation called convolution which works by taking a cluster of pixels and multiplying each pixel with the convolutional filter corresponding (the one that the neural network learn) and making a sum of the multiplyed pixels, repeating this process for all the pixels in a given image, the result will be a feature map. The feature map will be aplied to another test image and used like a feature identifier, a convolution operation will be made and for avery cluster of the test image the filter will be aplied, if the test image has the given caracteristics the result will be a high number ,making possible to identify, in this case, what is or what is not a brain tumor.
The maximum pooling layer is a compression layer, reducing the computation of all the convolution steps. The max pooling works by taking a cluster of a given size and shifting around the image and for each movement either taking the avarage of the values or choosing the max value of the cluster.
The database contains 4600 different images, with various sizes and shapes taken from Kaggle, with MRI images classified as healthy or unhealthy.
Comparing the accuracy in the training dataset with the accuracy in the validation dataset (by epoch)
