Abstract:
In terms of the neurological system, the brain is fundamental. A brain tumor is a growth or mass of abnormal tissue in the brain. And it's not uncommon for this to prove fatal. When a brain tumor is detected early, it can be treated more quickly and more people can survive. Brain tumors can be spotted on MR scans. However, there are occasions when that just isn't enough. When it comes to medicine, the ability to segment images is crucial. Which may be used to identify malignant brain tumors. However, there are numerous obstacles in the way of image segmentation. One such issue is the vanishing gradient. Which means more time and computing power may be needed to train Deep Convolutional Neural Networks. In order to address the vanishing gradient problem, we present a Deep Convolutional Neural Network (CNN) for fully automatic brain tumor segmentation in MRI data. The suggested method begins with a Resnet-50 classification of brain MRI images to determine the presence or absence of tumor. After that, we compared the two CNN models' degrees of accuracy. After that, we switched to using the U-net structure with the Resnet-50 encoder. And it's yielded spectacular results so far. Resents' skip connections, which serve as gradient superhighways, allow the gradient to travel unimpeded. This characteristic allows gradients to be propagated to higher layers before being attenuated to negligible or zero levels. Our model has been evaluated using the openly accessible LGG segmentation dataset. Our model is used to MRI scans only after they have been preprocessed and enhanced using techniques including rotation, zoom, horizontal and vertical shift, horizontal and vertical shear, and flipping. Results from our proposed technique have shown an accuracy of 99.7 %. As a benchmark, we've also used a model that combines U-net architecture with Inception V3, which has shown accuracy of 99.55%. Since our proposed model yielded better results, we were able to use it to identify and locate the tumor in the brain.
