Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 3

PREDICTION OF BRAIN TUMOR ON MRI IMAGES USING DEEPLEARNING

Method for tumor detection and multiclass brain tumor


classification where a contour-based method is used
The classification was made using ANN

The inclusion of more number of classes can degrade


this system.
Method for brain tumor detection and image segmentation

Image Segmentation- using adaptive thresholding and Canny edge detection


method
Brain Tumor Detection -Feature extraction was done using both the LOG-
Linderberg algorithm and Harris algorithm
Classifiers-Neural Network
proposed a paper on glioma detection based on multifractal features of segmented
brain MRI by Particle Swarm Optimization (PSO) technique

The SVM classifier used here is the nonlinear SVM

method which classifies normal brain and tumors such as glioblastoma, sarcoma and
metastatic bronchogenic carcinoma tumors
Classifier-DNN
Image Segmentation- Fuzzy c-means and DWT decomposition used for feature
extraction.
Note: DNN classifier gives good result in performance measures such as classification
accuracy, precision and AUC.
method for brain tumor detection
which involves CNN evolved through genetic algorithm
method for tumor classification using deeplearning algorithm.
Here faster R-CNN was used.Faster RCNN includes RPN(Region Proposal Network)
and RCNN

Among the various deeplearning techniques convolution neural networks are most
suitable for image processing applications due to its architecture and feature
extraction ability from 2D data like images.

Pre-processing operations are necessary on the raw image data as


the raw image does not give the accurate image details and can lead
to misclassification. The image dataset undergoes a number of pre-
processing steps such as gray scale conversion, blurring,
thresholding, erosion, dilation and finally cropping. Blurring is used
for image denoising. Gaussian blurring is used here. Binary
thresholding is used to convert the image. into binary image. Erosion
and dilation was done in two steps to eliminate the least amount of
unwanted region present and finally contour cropping was done on
images. The contour cropping was done by finding contours,
selecting the largest one ,find the extreme points and crop the
images

Data augmentation is a process which helps to create more images


from the available images itself by rotation, flipping and translation
procedures data augmentation can be applied as it increases the
accuracy of the system.
Convolution: The main purpose of the convolution step is the feature
extraction from the input mage. Here a filter is convolved with input
image pixel by pixel. The filter should be of the same depth as of the
input. The filter does the element wise multiplication with the
original values in the pixel. The multiplications are summed up which
results in a single number. The resulting array is called activation
map or feature map. The CNN can learn the values of the filter on its
own during training process.
Pooling: Pooling may help to reduce the number of parameters
required and the amount of computation required. It also helps to
control the problem of overfitting. Maxpooling is the most commonly
used one. Here the input image is partitioned into some sets of areas
that doesn’t overlap and the output of each area is the maximum
value in it. This reduces the size of the feature map with fewer
parameters.
Fully Connected Layer: This is a typical multilayer perceptron which
classifies the classes with greater accuracy. At this step the error is
calculated and then back propagated, the weights are adjusted so as
to optimize the performance of the model and this is how the
network trains the data.

Some of the popular CNN architecture types such as LeNet(1998),


AlexNet(2012), ZFNet (2013), Google Net (2014), VGGNet(2014),
ResNet( 2015). All these differs in the parameters used and in their
capability.

You might also like