Page 105 - The-5th-MCAIT2021-eProceeding

researchers proposed various image processing and deep learning methods for cancer detection, segmentation
and visualization in three-dimensional way.
A deep learning classification on MRI images for brain tumor detection and classification was proposed by
Rathi and Palani (2015). They performed tumor classification using multiple kernel based probabilistic
clustering and deep learning classifier. In the segmentation part, median filtering is used for image pre-
processing and multiple kernel based probabilistic clustering. In feature extraction module, shape, texture,
intensity based features are extracted. Linear Discriminant Analysis (LDA) was used for selecting important
features. Deep learning classifier is used in classification module that is employed having two important
processes of training phase and testing phase. However, their research requires further improvement towards
robust experimentation. Müller et al. (2016) proposed an automatic brain tumor segmentation with deep neural
network. Their proposed networks are tailored to glioblastomas pictured in MR images. To segment a brain,
their proposed method requires between 25 seconds and 3 minutes that is one order of magnitude faster than
most state-of-art methods. Convolutional Neural Network (CNN) was used to implement a novel two-pathway
architecture that learns about the local details about the brain. However, their research requires further
investigation towards lower computational complexity.
Damodharan and Raghavan (2015) proposed combined tissue segmentation and neural network for brain
tumor detection where pre-processing part consists of skull stripping. The skull removed MRI images are
employed for further classification of the brain tissues. In their research, initial steps involved in feature
extraction which aimed to find the neighbor blocks of the entire divided blocks, finding distance between all the
neighbor blocks, finding the feature values of the blocks with distinct distance measures. After the feature
extraction by Damodharan and Raghavan (2015), MRI image classification using neural network starts using
Feed Forward Neural Network (FFNN). However, Damodharan and Raghavan (2015) could concentrate more
in denoising issues to improve overall performance.

3. Proposed Conceptual Framework

Conceptual framework proposed by this research consists of two main parts, i.e. preprocessing and
segmentation part. Preprocessing part mainly contains noise reduction of MRI images and thresholding for
foreground extraction. In this context, this research intends to use median filter and adaptive thresholding to
denoise image in order to provide noise free image to the segmentation part. Another object of preprocessing is
to enhance contrast of the images. In the preprocessing phase, this research intends to use binarization
thresholding to extract whole head section in order to eradicate the background of the MRI images that may
interfere with further processing. After this, adaptive histogram equalization is intended to be used to enhance
the contrast of the images. However, in the context of MRI images, overall contrast is expected to be densely
distributed in some portion of the image, not throughout the whole image. So, adaptive histogram equalization
initializes a window for equalizing consisting pixels. Besides, this step is expected to provide clipping the excess
contrast and redistribute among all histograms bins (Arahusky, 2021).
After preprocessing, maximum noise free image will be inputted in the convolutional neural network (CNN)
where in this context this research is intended use a pretrained neural network. In the deep convolutional neural
network that is intended to be used by this research, encoder part will be consisted of two convolutional layers.
The first one is of stride 1 and the second one is of stride 2 followed by a max pool layer that uses nearest
neighbor method for down-sampling which will be repeated two more times, each for lower resolutions. The
decoder module will have other modules like encoder, i.e. max pool layer, an unpooling layer used for up-
sampling.

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [92]
Artificial Intelligence in the 4th Industrial Revolution