Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Neuroscience Informatics 2 (2022) 100062

Contents lists available at ScienceDirect

Neuroscience Informatics
journal homepage: www.elsevier.com/locate/neuri

Artificial Intelligence in Brain Informatics

Dynamic architecture based deep learning approach for glioblastoma


brain tumor survival prediction
Disha Sushant Wankhede ∗ , R. Selvarani
Dept. of Computer Science and Engineering, Alliance College of Engineering, University Campus, Anekal, Bengaluru, Karnataka 562107, India

a r t i c l e i n f o a b s t r a c t

Article history: A correct diagnosis of brain tumours is crucial to making an accurate treatment plan for patients with
Received 26 November 2021 the disease and allowing them to live a long and healthy life. Among a few clinical imaging modalities,
Received in revised form 21 February 2022 attractive reverberation imaging gives extra different data about the tissues. The use of MRI-Magnetic
Accepted 21 February 2022
Resonance Imaging tests is a significant method for identifying disorders throughout the human body.
Keywords:
Deep learning provides a solution for efficiently detecting Brain Tumour. The work has used MRI images
Glioblastoma for predicting the glioblastoma of brain tumours. Initially, data is retrieved from hospitals in form of
Magnetic resonance imaging an image database to continue with the brain tumour prediction. Pre-processing of dataset images is
Modified fuzzy C means a mandatory step to enhance the accuracy and smooth line supplementary stages. The intensity value
Rough set theory-based grey wolf of each MRI (Magnetic Resonance Imaging) is subtracted by the mean intensity value and standard
optimization deviation of the brain region. Further, reduce the medical image noise by employing a bilateral filter.
Overall survival prediction Further, the preprocessed medical images are used for extracting the radiomics features from images
as well as tumour segmentation. Thus the work adopts the tumor is automatically segmented into four
compartments using mutually exclusive rules using Modified Fuzzy C Means Clustering (MFCM). The
clustering-based approach is very beneficial in MR tumour segmentation; it categorizes the pixels using
certain radiomics features. The most important problem in the radiomics-based machine learning model
is the dimension of data. Moreover, using a GWO (Grey Wolf Optimizer) with rough set theory, we
propose a novel dimensionality reduction algorithm. This method is employed to find the significant
features from the extracted images and differentiate HG (high-grade) and LG (Low-grade) from GBM
while varying feature correlation limits were applied to remove redundant features. Finally, the article
proposed the dynamic architecture of Multilevel Layer modelling in Faster R-CNN (MLL-CNN) approach
based on feature weight factor and relative description model to build the selected features. This reduces
the overall computation and performs long-tailed classification. This results in the development of
CNN training performance more accurate. Results show that the general endurance expectation of GBM
cerebrum growth with more prominent exactness of about 95% with the decreased blunder rate to be
2.3%. In the calculation of similarity between segmented tissues and ground truth, different tools produce
correspondingly different predictions.
© 2022 The Author(s). Published by Elsevier Masson SAS. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction primary tumour [1]. In accordance with the World Health Organi-
zation (WHO), brain tumours are classified as benign (least dan-
Brain tumours develop when abnormal cells proliferate in the gerous) or malignant (more dangerous) based on the origin and
brain. Children and adults with brain tumours have a much higher behaviour of cells. Among adult patients with primary brain tu-
mortality rate than those without. Brain tumors are classified into mours, GBM, the most common and aggressive form of cancer, is
two major categories based on their malignancy and growth rate. the most common and most aggressive form. Even though GBM is
The first category is primary and the second, secondary. Primary treated with chemotherapy, radiation therapy, and maximum sur-
tumours emerge from the brain while secondary tumours form gical resection, the prognosis is poor [2]. While these multimodal
in other parts of the body when malignant cells grow from the therapies were unable to completely eliminate this dreadful illness,
therapeutic resistance and GBM recurrence were unavoidable. MRI
is commonly used in the diagnosis, prognosis analysis, and ther-
* Corresponding author. apy or other treatment planning of patients with GBM since it is
E-mail address: disha.wankhede@viit.ac.in (D.S. Wankhede). a non-invasive medical imaging technology. The picture is used by

https://doi.org/10.1016/j.neuri.2022.100062
2772-5286/© 2022 The Author(s). Published by Elsevier Masson SAS. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

MRI to extract compositional, structural, functional, and physiolog- of the data, small patches of the image surrounding each voxel are
ical information. MRI can create in vivo multidimensional images analyzed [14]. In addition, CNN architectures are used for segment-
of GBMs by recording this data, making it one of the most potent ing the tumour and extracting features such as shape, histogram,
methods for identifying the illness. MRI segments are mostly man- and geometric properties from it. Precise and hearty forecast of
ually annotated to determine brain tumour segmentation, which in general endurance through computerized calculations for pa-
impacts many treatment decisions, other treatment planning, and tients determined to have gliomas can give important direction
overall survival calculations [3]. Glioma patients need accurate to the conclusion, treatment arranging and result in expectation
prognoses to make informed decisions before undergoing treat- [13]. Researchers face the challenge of identifying tumors in low-
ment, and this forms the foundation for shared decision making complexity networks as one of their main challenges. According
[4]. to a recent study, meningiomas, gliomas, and pituitary tumors can
be classified accurately, but their approach requires a high number
1.1. Glioblastoma imaging modalities with its functional cases of convolution layers and kernels, which increases the computa-
tional cost. A second extension of the model identified the bound-
Checking procedures used these days for the brain growth ing box of the meningioma tumor. Although the paper focuses on
discovery, for example, X-beam, MRI examines, Computed Tomo- non-medical applications, it offers insight into the possibility of us-
graphic Scanning (CT filter) and the images nature got by those ing Faster-R CNN in medical imaging [30]. This article proposes a
strategies become better and better over the years [8]. MRI is the profound learning approach for Glioblastoma tumour survival pre-
most well-known strategy for brain growth determination to no- diction.
tice the harmful tissues and to find the influenced tissues in the Brain tumours are a central nervous system’s neoplasms het-
cerebrum which is successful for identification and grouping of erogeneous group nearby to the brain, that will arise. A definitive
various sorts and grades of cancers in the clinical analysis [9]. Fur- diagnosis and surgical therapeutic options were obtained, and the
ther, when a determination is set up, MRI is utilized to screen tumour location inside the brain has a thoughtful influence on
cancer all through the therapy worldview to evaluate the degree these patients’ symptoms. On the patient’s life quality, the brain
of careful resection, therapy reaction and infection movement, ac- tumour location also has a major influence on the neurological
cordingly fundamentally adding to GBM patient’s consideration toxicities risk, which can have a negative impact. By imaging after
[10]. Radiomic highlights removed from MR pictures utilizing pro- the beginning of neurological symptoms, brain tumours are cur-
gressed numerical calculations utilizes quantitative element extrac- rently only diagnosed. Due to their genetic makeup for exact kinds
tion and may uncover cancer qualities that neglect to be valued of brain tumours, even those individuals are recognized to be at
by the unassisted vision. These hand-tailored provisions are cho- risk and there are no early detection measures in place. For the
sen for either customary factual techniques or profound learning detection of brain tumours, many ML methods were used; many
calculations [11]. A deep learning-based radiomics model for pre- are supervised methods among them. Various researches follow
dicting survival for brain tumours and classifying their molecular the process such that the system fails due to coefficient overhead,
characteristics has been proposed [20]. In order to estimate over- overfitting and less effective feature extraction. The selection of
all survival in patients with glioblastoma, Nie developed a CNN- numerous layers in the learning and classification system is not
based radiomics model using a MRI set of 75 patients (training optimized in a deep learning method. As a result, the computing
dataset) and 37 patients (validation dataset). When six deep fea- overhead is increased. In addition, the classification process based
tures were selected from the pertained CNN, the established model on the whole MRI processing in each orientation results in overfit-
predicted overall survival better than conventional models [21]. A ting issues. Segmentation of images with a reference model is re-
unique advantage of these techniques is that they have the poten- stricted due to the stubborn edge regions. Classical ML techniques
tial to fundamentally change the way in which imaging is used to have a constraint of learning delay and accuracy of detection.
improve clinical care for patients with brain tumours. Such tech- Therefore, the need for an advanced deep learning approach arises
niques have their limitations, which is important to keep in mind, due to the time-consuming, inaccurate and erroneous system. This
and significant challenges remain. In the paper, CNN features were motivates the development of a new novel dynamic network-based
used in survival prediction to create a novel model for glioblastoma faster R-CNN modelling approach to predict glioblastoma cancer in
in relation to radiomics. In the context of automated learning and MRI tumour detection.
decision making, this helps to improve the accuracy of the predic- The remainder of the article is depicted as follows. Section 2
tion model. shows the survey of traditional writing dependent on brain cancer
characterization and endurance expectation. The proposed strat-
1.2. Glioblastoma brain tumour survival prediction egy details in section 3, experimentation and result discussion are
introduced in section 4 and conclusion of the work in section 5
Over the past decade, work in this area has turned to ma- respectively.
chine learning methods. The segmentation of brain tumours can
be considered from a machine learning perspective as a voxel- 2. Literature survey
level classification task to decide whether a given voxel belongs
to a normal brain, glioma, or edema class. Methods used to gen- The survival prediction of Glioblastoma brain tumours in MRI
erate such features can be broadly classified into two main cate- images requires step by step procedures for accurate results. Con-
gories: synthetic features and methods based on classic machine venient methodology selection is very much important for this
learning, such as support vector machines (SVMs) and random medical image processing. Various approaches exist to date for
forests [12]. Recently, deep learning has become popular in this brain tumour prediction and survival rate. Some of the written
field, particularly for segmentation, classification, and regression. works are as underneath:
Deep learning methods used for brain tumour segmentation can Naser et al. [16] created based on a pre-trained Vgg16 con-
be divided into three categories, depending on the network archi- volution, a transfer learning is generated for tumour segmenta-
tecture: patch-wise, semantically-based, and cascaded-based. There tion, U-net is presented based on CNN, and for tumour grading
are a lot of popular CNNs that use patch-wise techniques in solving an entirely associated classifier is used. For training and evalua-
segmentation problems today [13], Using a hierarchy of increas- tions, there were 110 patients with low-grade glioma (LGG), T1-
ingly complex and abstracted features of nonlinear representations postcontrast MRI images, FLAIR that is fluid-attenuated inversion

2
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

recovery, T1-precontrast pipeline used the grading and segmen- From MR images in glioblastoma multiform, overall survival is
tation models. The segmentation model achieved the tumour de- predicted by an approach proposed in Sanghani et al. [22]. From
tection accuracy and the Dice similarity coefficient (DSC) mean as 163 patients, they derived patient age, volumetric features, tu-
0.92 and 0.84, respectively. With accuracy, specificity, and sensitiv- mour shape, and textural features. Both 2-class (short and long)
ity of 0.89, 0.92, and 0.87 at the level of MRI images, and at the and 3-class (short, medium, and long) survival groups had their OS
patient level the values are 0.95, 0.97, and 0.98, grade II and III are groups predicted. For feature selection, a recursive feature removal
the two grades under the grading model which is divided by the using the SVM model was applied. The 5-fold cross-validation
LGG. is used to evaluate the classification model’s performance. The
The breast cancer’s diagnosis prediction is improved by the pro- 88.95% is the obtained accuracy for the OS group’s predictions. For
duced informative features and employed multi-modal data, for the first time, to estimate the GBM patient’s prognosis, the shape
this with the random forest classifiers, gated attentive DL mod- characteristics were used. Both 2-class and 3-class OS group pre-
els loaded were developed by Arya et al. [17]. The stacked features dictions were having higher accuracy for the feature selection and
are generated by the sigmoid gated attention CNN is the phase one prediction scheme.
and to the random forest classifier, the stacked features are passes Glioma patients risk calculation and survival prediction with
is the phase two, these are the bi-phase model of this study. In the existence associated to 7 gene signatures by the proposed
the survival estimation of breast cancer patient’s sensitivity value two techniques of Wijethilake et al. [23]. They suggested a novel
was found to be 5.1% in a comparative analysis with different ap- probabilistic programming-based approach that outperforms the
proaches. Extract tumour genotype-related features and feed them existing classical ML algorithms for survival prediction. The rec-
into OS prediction from the pre-operative MRI brain data, where ommended technique attains a 4-fold accuracy of 74% on average.
a novel DL-based Overall Survival (OS) prediction technique is in- Benjamin et al. [7], Karayegen et al. [15], Baid et al. [24] proposed
troduced by Tang et al. [18] for GBM patients. To perform both OS using T1ce MRI images and FLAIR to extract first-order, intensity-
prediction and tumour genotyping tasks simultaneously, a multi- based volume, shape-based, and textural radiomic characteristics
task CNN was created. Predicting OS time accuracy is significantly for glioblastoma survival prediction. With the low and high pass
improved by learning tumour genotype-related characteristics for filtering based on the stationary wavelet transform, the area of in-
genotype prediction. Four separate genotypic molecular indicators terest was further decomposed. The acquisition of directional infor-
mation is aided by the extracted radiomic on those deconstructed
were used in the experimental design, which used 120 GBM pa-
images. Testing, validation, and training datasets of Brain Tumour
tients’ multimodal MRI brain dataset.
Segmentation (BraTS) challenge, the algorithm’s efficiency was as-
Amjad Rehman et al. [19] introduced a deep learning-based
sessed. Training, validation, and testing of the BraTS, the technique
approach to detect and characterize microscopic brain tumours.
attained 0.695, 0.571, and 0.558, respectively.
The first step consists of designing a 3D convolutional neural
Anand et al. [25] used preoperative MR images to examine the
network (CNN) architecture, which is used to extract brain tu-
efficacy of a U-net neural network to correctly identify three lo-
mours and train a pretrained CNN model to extract features. Using
cations of a brain tumour and an ELM to forecast patient overall
the correlation-based selection method, the extracted features are
survival after gross tumour excision. For training, 210 GBM pa-
transferred to this method, and as an output, the best features are
tients were used, and for validation, 66 GBM patients and Low-
selected. Final classification is based on the feed-forward neural
Grade Glioma were employed. Before loading the data into the
network that validates the selected features. Experimental and val-
model, each patient’s data went through a sequence of prepro-
idation studies were conducted with three BraTS datasets (2015,
cessing processes. One for each region of interest, likewise three
2017, and 2018), and the accuracy was 98.32, 96.97, and 92.67%
separate U-nets make up the segmentation model. These segmen-
respectively.
tations were then compared to physician-created contours using
Suter et al. [20] described the machine learning (ML) method
established quantitative criteria. The 59 high-grade glioma patients
by classifying the patient’s MRI image for automatically predicting with gross total resection (GTR) were used to train on patient over-
the patient’s survival rate with glioma brain tumours. The patient’s all survival prediction.
age and their overall survival duration in days, and each sample in- Goriparthi et al. [5] investigate and implement the CNN archi-
cluded four MRI brain imaging sequences, where 163 samples are tecture with a three class classification system, 98% of the training
collected from the BraTS 2017. Long-term, mid-term, and short- and 96% of the testing accuracy are achieved.
term are the three categories of survivors. Various ML algorithms The goal of the study by Kociołek et al. [26] was to see how
trained and extracted the histograms and deep features, intensity, alternative images normalising methods and the number of inten-
statistical, and volumetric texture to improve prediction results sity levels affected texture classification, while also accounting for
these are the characteristics of many features. Linear discriminant, noise and artefacts caused by uneven backdrop brightness.
k-nearest neighbours (KNN), SVM, logistic regression, tree, and en- Sathiyabhama et al. [29] proposed a hybridization of GWO and
semble were among the ML approaches used. Rough Set (GWORS) approach for relevant characteristics from
Wijethilake et al. [21] investigated the impact of integrating the retrieved mammography images also compare the proposed
genomic and radiomic characteristics on the overall survival (OS) GWORS to other well-known rough set and bio-inspired feature
prognosis in GBM patients. To estimate survival rates for each pa- selection methods, such as particle swarm optimise, genetic algo-
tient, they used a hyper column-based convolutional network to rithm, Quick Reduct, and Relative Reduct, to see how successful it
separate tumour regions from MRI, extract radiomic properties is.
(histogram, shape, and geometric) and combine them with data S.L. Bangare et al. [36] discussed the experiments with applica-
from gene expression profiles. To conduct prognosis analysis, sev- tions of machine learning in medical domain. N. Shelke et al. [37]
eral state-of-the-art regression models were used, including neural have worked on the LRA DNN methods.
network, linear regression, and support vector machine (SVM). The The existing approaches reviewed regarding the processing of
model’s function in genomic, radiomic, and radio-genomic aspects MRI images for Glioblastoma survival prediction possess many se-
was observed using the gene expression and MRI profiling dataset rious issues such as uneven images, inaccurate prediction results
as Cancer Genome Atlas (TCGA). Both radiomic and genomic mod- with time complexity for processing. To overcome those mentioned
els are outperformed by the radio-genomic model and the GBM OS issues, a deep learning-based approach for glioblastoma brain tu-
prediction is associated with genetic data is revealed in the results. mour survival prediction has to be proposed.

3
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Fig. 1. Workflow of the Proposed Methodology.

3. Proposed research methodology feature segmentation. The segmented images may retrieve enor-
mous features. The most significant and informative features from
Brain tumours are difficult to be modelled with normal tis- the extracted features were selected utilizing Rough Set Theory-
sue or significant size overlapping intensities, forms result. More based Grey Wolf Optimization. Then, using an FR-CNN, the clas-
accurate diagnosis is made, the MRI brain image data advan- sification for total survival estimation is done to feature selected
tages, the software-based medical image processing are numerous. images. Thus, the classification technique increases the accuracy
In machine learning, computer-assisted tumour detection systems with lesser learning time and improved convergence rate. The pro-
and CNN have shown to be successful and have made significant cedures that must be followed in the proposed glioblastoma brain
advances. This section describes the deep learning approach for tumour survival prediction system are as follows.
Glioblastoma brain tumour survival prediction. MRI brain tumour
images were used for the prediction of glioblastoma. Fig. 1 por- 3.1. Data collection
trayed the workflow diagram for the suggested methodology.
The proposed approach comprises four stages. In the first step, The dataset has been collected from the Kaggle dataset with
the data for usage is collected from a static weblog and pre- an entire number of 253 MRI brain tumour images. From each set
processed to eliminate all of the contaminants from the data. of images, the entire process of survival a forecast has been made.
Meanwhile, the image contains Gaussian noise, salt pepper noise, During pre-processing, the MRI brain pictures are obtained and fed
and so on, which are the different forms of noise. Addition- into it as input.
ally, images are captured with separate scanners in a few situa-
tions, and image intensities are utilized to normalize the images, 3.2. Data pre-processing for noise and bias reduction
thus the pre-processing comprises intensity normalization by his-
togram normalization and de-noising utilizing bilateral filter has Enhance the image data based on the preprocessing step and
been done to enhance the images. The principal objective in pre- for subsequent processing, the image attributes improvement is
processing is to develop the dataset by settling the noisy infor- important. Gaussian noise, salt pepper noise, and so forth are pre-
mation and recovering the missing qualities. Next, segmentation sented in the MR images. The intensities had to be normalized as
is carried out using the MFCM clustering algorithm for radiomic well because the images were captured with different scanners.

4
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Images are taken using different scanners, so intensities also are maximum (HIR and T 2l ), and mean (πi and πs )], will be easily
normalized. This article utilizes Histogram Normalization and Bi- reached without relying on incorrect landmarks in the histogram
lateral Filtering for reducing noises and biases in the image. are discussed in this article. As a result, the suggested normaliz-
ing method differs from the existing normalization method. Using
3.2.1. Histogram normalization for intensity normalization in images noise estimation, the performance of histogram normalization is
In brain MRI analysis, the crucial preprocessing step is inten- measured after normalizing the input image to the reference im-
sity normalization. During MR image collecting, various scanners age.
or parameters may be used for scanning different people or the
same topic at different times, resulting in significant intensity vari- 3.2.2. Image de-noising using bilateral filter
ances. This intensity fluctuation will substantially hinder future The MRI glioblastoma images may contain noises such as Gaus-
MRI processing and population analysis, such as image registration, sian, salt pepper noise etc. Removing the noise preserves the data
segmentation, and tissue volume estimation. To improve image on the same as in the input data. A bilateral filter is used for de-
similarity and make MR image comparisons between MRI scans noising these input images. Without using the smoothing edges,
easier, a normalization algorithm changes the distributions of each the spatial weighted averaging is applied by the bilateral filter [27].
follow-up scan to match the chosen baseline scan. Histogram nor- By merging two Gaussian filters, this is accomplished, in the do-
malization algorithm has been utilized without requiring any phys- main of spatial one of which operates and in the intensity domain
ical intervention or prior knowledge for intensity normalization. the other one is operating. For the weight, both the intensity and
Using various acquisition parameters and differing field
spatial distance are used. The bilateral filter output at pixel loca-
strengths were similar for the same person that for the same
tion p can be described as follows:
intensity type proportions of intensity levels for the same tissue
1 
type in brain MRIs from scanners. As a result, for the input im- −q− p 2 −| F (q)− F ( p )|2
2εe2 2εs2
age, a low-quality image will be used and for the reference image, F ( p) = e e F (q) (3)
N
a high-quality image will be used. Frequently the issues can be z∈ S ( p )
caused by the histogram’s tails. Resulting in significant scanner
Where, S ( p ) is a pixel’s spatial neighbourhood F ( p ), and N is
fluctuations, the tail of high intensity usually corresponds to arte-
the normalization constant, εe and εr are parameters governing
facts and outlier intensities. To prevent this issue, the reference
image is first preprocessed by removing the background and out- weights in the domains of intensity and spatial start to fall off.
liers, resulting in the Intensity of Interest (IOI) being used as the  −q− p 2 −| F (q)− F ( p )|2
2εe2 2εs2
standard scale. The following are the two steps that make up the N= e e (4)
overall strategy. z∈ S ( p )
The T max , T min are the maximum and on the standard scale,
Bilateral filters have been employed in tone mapping, volu-
minimum intensities (matching to the standard histogram). The
metric de-noising, texture removal, and other applications like
Low-Intensity Homogenous Regions of Interest (LIR) and High-
de-noising the image. To the dimensions of the original domain,
Intensity Interest Regions (HIR) are used to make up the histogram
in a higher-dimensional space that express the filter, the inten-
for the reference image. At the LIR, the histogram was started sity for the signal is added. They may construct simple criteria
and goes all the way up to HIR. To the HIR and LIR values, the for down-sampling the critical processes and achieving accelera-
image intensities are mapped. To encompass all of the input im- tion by expressing in this augmented space, by the two modest
age’s grayscale levels, the reference image histogram is shifted and nonlinearities, the bilateral filter is implemented as simple linear
stretched as shown below. convolutions.
HIR − LIR
h (a, b, c ) = (h (a, b, c ) − T min ) + LIR (1) 3.3. Tumour region segmentation
T max − T min
In IOI, at T min the input image’s target histogram h (a, b, c ) Using a range of MR imaging techniques such as diffusion MRI,
starts and spreads up to T max grayscale levels, scaled up between T1, T2, and proton density imaging, brain tumour segmentation de-
the image of the lower boundary n1 and n2 is the upper boundary, tects the position and extent of aberrant tumour areas, as well as
and between the minimum (LIR) level and a maximum (HIR), the whether the tumour region is vascularized, swollen, or infiltrative
new normalized image’s voxels h (a, b, c ) will lies. The reference Comparing the images from various imaging techniques can iden-
image’s upper and lower bounds before scaling up are represented tify which types of tumours show up clearly in certain modalities.
In the first phase of the algorithm, a tumour is automatically seg-
by the variables n2 and n1 . Through the use of two different linear
mented into four compartments that are mutually exclusive: (C1)
mappings, this is accomplished. From [T 1l , πl ] to [LIR, πs ], the first
necrosis, (C2) enhancing tumour, (C3) T1 abnormality excluding C1
is ranges while from [πl , T 2l ] to [πs , HIR] the second ranges. Let
and C2, and (C4) FLAIR abnormality excluding T1 abnormality. Each
the normalization function is denoted as O (a, b, c ). The expression
voxel was classified into one of four tumour compartments us-
for O (a, b, c ) is: ing a modified fuzzy c-means clustering method (or as normal).
⎧  The three imaging sequences that are available suggest, the clas-

⎪ πs + (h (a, b, c ) − πi ) LIR−πs
,

⎪ T 1l −πs sifier generated 219 low-level characteristics for each voxel (the

n1 ≤ h(a, b, c ) ≤ πi  sequences were registered prior to segmentation).
O (a, b, c ) = (2)

⎪ πs + (h (a, b, c ) − πi ) HIR−πs
,

⎪ T 2l −πs
⎩ 3.3.1. Modified fuzzy C means clustering algorithm
πi ≤ h(a, b, c ) ≤ n1 The fuzzy memberships used by the  FCM method are used to
Where the input image histogram and reference image his- assign pixels to each category. Let X = xi , i = 1, 2, . . . , N |xi ∈ R d
tograms mean values were represented as πi and πs , and the signify an N-pixel picture that will be partitioned into c classes
function of the ceiling is referred to as ·. The input image’s (clusters), with xi representing feature data. The method is an it-
voxel values are represented as T 1l and T 2l . For normalizing the erative optimization process that seeks to minimise the objective
MRI images, three crucial intensity values [minimum (LIR and T 1l ), function denoted as

5
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062



c 
N
min P (γ ( S )) ds (9)
Jm = um
ki xi − vk 2
(5) γ
k =1 i =1 γ

Further, this is objective function defined in (5) was immedi- Where, γ is a route that connects v i to v s . The optimization
ately adjusted as follows: was done using the fast marching method, with the weight P at
each voxel proportional to the gradient magnitude. A grey-level

c 
N
α 
c N  co-occurrence matrix provides texture statistics for the first and
Jm = um 2
ki xi − v k  + um
ki xr − v k 2 (6) second order. The first-order statistics for each voxel are the av-
Nr
k =1 i =1 k =1 i =1 r∈N i erage and variance of the intensities within two voxels of each
modality. The picture volumes were first normalized to 64 distinct
Where, N i is the cardinality of a group of neighbours falling in-
grey levels for the second-order statistics, and then every image
side a local window xi and N r . The penalty’s effect is controlled
slice was bound by a bounding box consisting of 5 × 5 pixels.
by the second term’s parameter alpha. In essence, adding the sec-
Finally, by examining based on a probabilistic model, revis-
ond term to (6) creates a spatial restriction that seeks to maintain
ing the spatial configuration of the current segmentation labels,
continuity on neighbouring pixel values around xi . The objective
a patient-by-patient refinement was done. To begin, the inten-
function J m can be minimised under a constraint by using an op-
sity distributions of voxels with a GLISTR posterior probability
timization method.
of 1 were populated individually for the tissue classes of white
In MRI scans, nearby pixels/voxels with comparable feature val-
matter, oedema, necrosis, non-enhancing, and enhancing tumour.
ues are called neighbouring pixels/voxels. The membership func-
It’s worth noting that the non-enhancing and necrotic regions
tion is adjusted in order to acquire the spatial information of the of the tumour aren’t distinguished in the present segmentation
pixels accurately. Its current definition is as follows: aim. The pair-wise distributions’ histograms were then normal-
y ized. The class-conditional probability densities (Pr (Ivi ) |Class1 )
v i j .sm
v i j =
ij
(7) and (Pr (Ivi ) |Class2 ) were simulated by fitting several Gaussian
n y
k =1 v kj .sm
kj
models and calculating the mean and standard deviation for each
 class using Maximum Likelihood Estimation. The oedema voxels
si j = v ik (8) opposed to the white matter voxels in the T2-FLAIR volume, the
k∈ N a j ET voxels opposed to the oedema voxels in the T1-CE volume,
and the ET voxels opposed to the union of the necrosis and the
Where, si j is a spatial function that represents the likelihood non-enhancing tumour in the T1-CE volume are the three pair-
of a pixel/voxel a j belonging to the i-th cluster. N a j is a square wise distributions studied here. The former intensity population is
window in the spatial domain that is centred on a pixel/voxel a j predicted to have substantially higher (i.e., brighter) values in all
in the spatial domain and y , m are parameters of the window. The circumstances. As a result, the intensity of voxels in each tissue
creative membership value is improved by the spatial function, and class that were in close spatial proximity (i.e. 3 voxels) to the op-
residues are unaffected by clustering. The stages of the MFCM Al- posing tissue class was examined.
gorithm are as follows: Specifically, the intensity I ( v i ) of each of these voxels was mea-
sured and (Pr (Ivi ) |Class1 ) was compared with (Pr (Ivi ) |Class2 ). Fol-
Step 1: Set the number of clusters c and the parameter
 y. lowing that the voxel v i was assigned to a tissue class based on
Step 2: Compute membership matrix V b = v i jb using Equa- the bigger of the two conditional probabilities.
tion (8),
Step 3: Compute t j using Equation (5), 3.4. Radiomic based features selection
Step 4: Repeat steps 2 and 3 untilthe resulting termination condi-
tion is satisfied:  V b − V (b−1)  < ε where b denotes the num- The image radiomic features such as GLISTR posterior proba-
ber of iteration steps (b = 0, 1, 2, . . .). bility maps, geodesic information, intensity information, texture
features, and image derivative segmented from the preprocessed
According to this segmentation model further, extract the image may suffer from overfitting problems and expand computa-
radiomic feature from the image. Intensity information, image tional complexity due to the large dimension of data. Moreover, a
derivative, geodesic information, texture features, and GLISTR pos- data dimensionality is reduced to find the significant features by
terior probability maps were among the radiomics features. These applying Rough Set Theory-based GWO (RS-GWO).
features have been utilized to develop diagnostic, prognostic, and
treatment response prediction models. To improve reliability, these 3.4.1. Rough set theory
models are being integrated with clinical, biological, genetic, and A few basic concepts are introduced in this section on rough set
proteomic characteristics. theory and RS-based feature selection. Let J = ( P , T , C , G ) be an
information system, a finite non-empty instance set is represented
3.3.2. Radiomics features extraction as P , a finite non-empty attributes set is denoted as T , and the set
The intensity component includes each voxel’s raw intensity union of attribute scopes are described as C , where C = U t ∈ T C t
value I ( v i ), as well as their differences across all four modali- denotes the value scope of attribute t . g : P × T → C . Each char-
ties (i.e., T1, T1-CE, T2, T2-FLAIR). The picture gradient magnitude acteristic has a unique magnitude that is associated with each
and the Laplacian of Gaussian make up the image derivative com- instance of the information function in P , resulting in g ( y , t ) ∈ C t
ponent. Prior to generating any intensity-based feature, intensity for any t ∈ T and y ∈ P . There is an indiscernibility connection
normalization was conducted using the GLISTR segmented cere- IR ( A ) linked with each P ∈ T . The A-lower and A-upper approx-
brospinal fluid’s median intensity value. At voxel vi, the geodesic imations of Y are determined as A ∗ (Y ) and A ∗ (Y ) in equations
information was provided by the geodesic distance from the seed- (13) and (14), respectively, for subset Y ∈ P and equivalence rela-
point utilized as the tumour centre in GLISTR, which was at voxel tionship IR ( A ).
vs. The geodesic distance between vi and vs was calculated as fol-
lows: IR ( A ) = {(a, b) ∈ P × P | ∀t ∈ q, g ( y , t ) = g ( z, t )} (10)

6
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

The P /IP ( A ) denotes the partition of P determined by IR ( A ) and where X  α is the position of the alpha, X β is the position of the
can be calculated as beta, X  1,
 δ is the position of the delta, C 1 , C 2 , and C 3 , and A
 2 and, A
A  3 are all random vectors, is the position of the current
P /IP ( A ) = ⊗ { T ∈ A : P /IP ({t })} (11) solution, and indicates the number of iterations. The purpose of
S ⊗ L = {Y ∩ z : ∀ R ∈ Y , ∀ L ∈ Z , Y ∩ Z = ϕ } (12) mathematical equations (22) to (25) is to mathematically model
circling behaviour, and they are used as follows:
A ∗ (Y ) = { y ∈ P : [Y ] A ∈ y } (13)

→ −
→ −
→−→

A (Y ) = { y ∈ P : [Y ] A ∩ Y = ϕ } (14) T ( u + 1) = T p ( u ) + B . E (22)


The equations (15) and (16) are used to identify the F -positive Where, E is characterized in equation (22) and u is the iter-
and F -negative regions of X for two subsets F , X ∈ T that give

→ −
→ −
→ −

ation numbers, B , D are vector coefficient, T p and T are the
growth to two equivalence relationships IR ( F ) and IP ( X ). As seen −
→ − →
praise and grey wolf position. The B , D vectors are computed in
in the equation, the F-boundary area of X is defined as follows: equations (20) and (21).
(17).

→ − 
→ −
→ −
→ 
E =  D . T p (u ) − T (u ) (23)
POS F ( X ) = ∪Y ∈ P /IND( X ) F ∗ (Y ) (15)

→ −
→−→ −

NEG F ( X ) = P − ∪Y ∈ P /IND( X ) F ∗ (Y ) (16) B = 2 b . s1 − b (24)
∗ −
→ −

Bd F ( X ) = ∪Y ∈ P /IND( X ) F (Y ) − ∪Y ∈ P /IND( X ) F ∗ (Y ) (17) D = 2 s2 (25)

Finding dependencies between attributes is an important prob- Where, s1 , s2 are vectors with random values in [0, 1], and the


lem in data analysis. If and only if IR ( F ) ∈ IR ( X ), it depends on component of b decreases linearly from 2 to 0 throughout it-
X dependencies that can be determined in the same way that F , erations. The alpha is usually in charge of the chase. From time
X  T , F . In the equation (18), the dependence degree is expressed, to time, the beta and delta may also participate in chasing. Al-
if X is dependent on F to the degree 0 ≤ μ f ( X ) ≤ 1. For attribute pha is the best candidate solution, second-best candidate solution
selection, in greedy algorithms the dependence degree μ f ( X ) can (beta), and third-best candidate alternative (delta) are predicted to
be utilized as a heuristic. have improved learning about the probable prey region to recre-
|POS F ( F )| ate it mathematically, the chasing behaviour of grey wolves. To
μ f (X) = (18)
|P | reposition themselves in agreement with the best search agents’
μd ( E ) = μsel ( E ) and ∀ Q ∈ Sel, μq ( E ) = μd ( E ) (19) positions, the first three best solutions obtained thus far compel
the other search agents (including the omegas). The wolves’ posi-
If μ f ( X ) = 0, X is completely independent of F , X is somewhat tions are then updated as depicted in Equations.
dependent on F, μ f ( X ) = 1, X is dependent on Q ; if 0 ≤ μ f ( X ) ≤ −

→ → −
→ → −
− → −
→ −
→ → −
− → −
→ −
→ →

1. Where E is the decision attributes set and D is the condition E ∝ =  D 1 . t ∝ − Y  , E β =  D 2 . t β − Y  , E δ =  D 3 . t δ − Y  (26)
attributes set if T = D ∪ E and D ∩ E = 0 in an information table, 
J = ( P , D ∪ E , C , G ) is named a decision table. The precision of −
→ −→ −
→−→ −


−→ −
→−→ −


−→ −
→−→
t 1 =  t ∝ − B 1 . E ∝  , t 2 =  t β − B 2 . E β  , t 3 =  t δ − B 3 . E δ  (27)
selection caused by the decision attributes set is denoted μd ( E ),

→ −
→ −

which is the dependency degree between condition and decision −
→ t1 + t2 + t3
attributes. To eliminate redundant attributes, a reduced set can still t ( u + 1) = (28)
3
be classified accurately as the original settings using the attribute


purpose reduction. In equation (17), an attribute that represents a The updating of the parameter b , which governs the balance
subset Sel of the condition attribute. between exploitation and exploration, is a final note regarding the


GWO. According to the equation, the parameter b is linearly up-
3.4.2. Grey wolf optimization
dated in each iteration, ranging from 2 to 0, with u being the
Grey wolves’ hunting behaviour and social leadership in nature
iteration numbers and mi being the total iterations maximum per-
inspired the GWO algorithm. The GWO method works similarly to
mitted for the optimization (25).
other metaheuristics in that it starts with a population of randomly
produced wolves (candidate solutions). The population of wolves −
→ 2
b = 2 − u. (29)
is divided into four groups in this algorithm to establish the so- mi
cial hierarchy of wolves while developing GWO: alpha, beta, delta,
and omega. The fittest answer is known as alpha (α ), while beta 3.4.2.1. Fitness function calculation for dimensionality reduction In it-
(β ) and delta (δ ) are the second and third most effective options, erations, until it finds a satisfying solution, the GWO explores new
respectively, within the GWO’s mathematical model. Omega (ω ) is regions in the attribute space and exploits solutions. The wolves’
believed to be whatever semblance of a hopeful solution. By these locations reflect attribute set selections and the solution space is
three candidates, the hunters are led by α , β , and δ and the ω made by all potential attribute selections for the RS combined
are saved. The mathematical model of readjusting the positions of GWO. The fitness function, as exposed in equation (30), is used
wolves is presented as follows. to determine whether an attribute subset will be selected or not.
    |D − S|
   
 α = C 1 · X
D α − X
,  β = C 2 · X
D  β − X  , Fitness =∝∗ γ S ( E ) + β ∗ (30)
  |D|
 δ = C 3 · X
D δ − X  (20) Where, | S | is the chosen attributes length subset, and γ S ( E )
    is the classification quality of condition attribute set S relative
 1 = X α − A
X 1 · Dα , 2 = X
X 2 · D
β − A β , to decision E. The total amount of qualities are indicated by the
  letter | D |. ∝∈ 0, 1 and β = 1− ∝, are two corresponding values
 3 = X δ − A
X 3 · D δ (21) for the relevance attribute subset length and classification quality.

7
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Fig. 2. Faster-RCNN Network Model.

Attribute subset length and quality of classification have distinct box regression offsets, the RPN network calculates proposals. This
implications for the attribute reduction task, according to this for- region proposal network accepts a backbone layer-derived convo-
mula. The attribute subset length and set ∝= 0.9, β = 0.1 are less lution feature map as input and outputs anchors formed by sliding
essential in this article than the quality classification. The high en- window convolution applied to the feature map.
sures that the ideal position is a rough set reduction at the very RoI Pooling: Based on the feature maps extract the proposal
least. This fitness function assesses the quality of each position. into the whole connection layer that follows using this layer. This
The fitness level is to maximize is the goal of the study. After de- layer receives the output of the region proposal as input and passes
termining the fitness level of each image, significant features are it to the ROI pooling layer, which performs the same function as
taken from the extracted images, and redundant features are re- Fast R-CNN in converting variable sizes of RPN region proposals
moved. into a fixed-size feature map.
Softmax and Bounding Box Regression: The feature map size
3.4.3. Tumour region classification using faster-RCNN obtained by RoI pooling is then transmitted to two fully connected
The features selected from RS-GWO are classified for predicting layers, which flatten the feature maps before sending the output to
the overall survival rate by FR-CNN. To extract features, for each two parallel fully connected layers, each with a separate job given
item suggestion, the R-CNN necessitates a forward pass through to it:
the convolutional network, resulting in a significant computing The loss function of FRCNN is denoted as in equation (31), in
overhead. Faster R-CNN is used to mitigate this problem. For the which q j is anchor j’s projected probability of being a thing. u j =
RPN and Faster R-CNN detectors, until they reach their full po-
{ua , u b , u c , ud } is a vector that contains the prediction’s bounding
tential, the key concept is to use the same convolutional layers.
box’s four parameterized coordinates. The ground truth bounding
To generate and refine object recommendations, the image is now
box by the positive anchor has a coordinate vector u ∗j . The ground-
only run through the CNN once. Using the Rol pooling layer, to
a fixed-size feature map each proposal is converted and the pro- truth label is mentioned in equation (32). Equation (33) depicts the
posal is mapped to CNN’s final convolution layer. Finally, to obtain logarithmic loss target and non-target as M cls (q j , q∗j ).
the needed carrier component, apply Softmax classification and
bounding box regression. The overall structure of Faster R-CNN is   1 
shown in Fig. 2. Proposed Faster R-CNN is used for brain tumour M qj , uj = M cls (q j , q∗j )
O cls
detection based on a deep convolution network. j
This work uses a more efficient Faster R-CNN deep learning 1  ∗
+δ q j M reg (q j , q∗j ) (31)
model to identify brain tumours. Convolution layer, region proposal N reg
j
network (RPN), and bounding box prediction are three phases
 
of the faster R-CNN tumour detection algorithm. The convolution 0, negative label
layer contains a filter that extracts appropriate image features, fol- q∗j = (32)
1, positive label
lowed by a small network of RPNs sliding over the convolution  
layer that consists of a feature map to predict whether or not the M cls q j , q∗j = − log[q∗j q j + (1 − q∗j )(1 − q j )] (33)
tumour is detected, and if the tumour is predicted, the bounding
box is used to detect the object by fully connected neural net-
M (u j , u ∗j ) is the regression loss, calculated using M reg (u j , u ∗j ) =
works. To improve tumour identification accuracy, we replace the
R (u j − u ∗j ), where R is the smooth M 1 function. While the other
convolution layer with a feature extractor network model, such as
ResNet50, in the general flow of traditional Faster R-CNN extra. anchors there is no (q∗j = 0), and q∗j M reg denotes that only the
This technique is known as Enhanced Faster R-CNN extra. foreground anchor(q∗j = 1) has a regression loss. The N cls and N reg
Convolutional Layers: Including the 13 convolutional layers are used to normalize and the cls and reg outputs are made up of
combined with 13 Relu layers and to extract image feature maps, {q j } and {u j }, respectively. The normalized results are the classified
four pooling layers were used. Following RPN layers and complete outcome of the deep learning approach. This dynamic computation
connectivity layers use the same feature maps. using a deep learning approach reduces the overall computation
Region Proposal Networks (RPN): The region suggestions are and perform long-tailed classification. With the development of
generated by this method. After passing through a 3 × 3 con- this Faster RCNN, the training performance of classification will be
volutional layer that generates foreground anchors and bounding accurate.

8
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

3.4.4. Dynamic architecture of multilevel layer modelling for survival Table 1


prediction Simulation System Configuration.
Dynamic networks can modify their structures, parameters, or Python Jupiter Version 3.8.0
pick salient spatial/temporal areas in the input based on data. It is Operating System Windows 11 Pro
normal to execute inference using dynamic structures conditioned Memory Capacity 6GB DDR3 RAM
on each sample, given that various inputs may have different pro- Processor Intel i5 @ 3.5 GHz
cessing needs. Within a Multi-Scale Dense Network (MSDnet) that
comprises numerous viable pathways, one may alter the network
working platform of PYTHON and the system configurations are as
depth, width, and perform dynamic routing. Networks with dy-
follows.
namic architectures not only avoid unnecessary computation for
Table 1 explains the working progress of proposed methods in
canonical (“easy”) samples, but they also keep their representation
the simulation environment of Python Jupiter Version 3.8.0 with
capacity while identifying non-canonical (“difficult”) samples. Us-
the system configuration of Intel Core i5 operating at the maxi-
ing 10-fold cross-validation, the researchers discovered that train-
mum clock speed of 3.5 GHz with a dedicated memory of 6 GB
ing MSDNet on the overall survival data of 253 patients resulted
of DDR3 class ram on the operating system of Windows 10 Home.
with increased accuracy for malignancy classification on a differ-
The simulation runs for almost 10 to 15 minutes with a maximum
ent tumour.
iteration of 100 to achieve the required result.

3.4.4.1. Multi-level layer of Multi-Scale Dense Network (MSDnet) Deep


4.1. Kaggle dataset description
Neural Networks (DNN) are becoming progressively deep for de-
tecting more “hard” samples, doing inference with dynamic depth The dataset for the proposed system’s implementation is taken
is a simple way to reduce unnecessary processing. The coarse-level from the Kaggle dataset. It consists of 253 MRI brain tumour im-
information required for categorization is missing from the high- ages. This dataset is about brain cancer gene expression. Brain
resolution features, resulting in unsatisfactory early exit findings. cancers are malignant brain tumours which is the growth of ab-
To address this problem, this multi-scale dense network (MSDNet) normal cells that have formed in the brain. Predictions of survival
makes use of 1) a multi-scale architecture consisting of numer- have been made for each set of input images. The 91 (70%) in-
ous sub-networks for generating coarse-level features suitable for stances are randomly chosen by the training data. The test data is
classification based on feature maps of varying resolution and di- the remaining 39 (30%) instances.
mensions; and 2) dense connections to reuse early features and
improve deep classifier performance. The total accuracy of the net- 4.2. Simulation output
work’s classifiers is effectively improved by such a particularly
built architecture. The suggested algorithm’s performance is evaluated using real-
As part of MSDNets, the classifiers rely on dense connectivity time images of tumour and non-tumor individuals. The results of
patterns within coarse-scale networks coarsest scale, S,  i.e., the the approach in conjunction with MSDNet-based Faster R-CNN,
classifier at layer l employs all of the features x1s , . . . xls . In the MFCM, are compared to traditional systems to assess its efficiency.
anytime mode, the model propagates the input through the net- Two separate real-time MRI images for recognising anomalies in
work until the budget is depleted, at which point it outputs the brain tumour images are acquired to measure the performance of
most recent forecast. An example traverses the network and exits the suggested hybrid technique.
after the classifier f k in the batch budget setting at test time if its Fig. 3 depicts a sample image from the dataset. The Kaggle web-
prediction confidence (the softmax probability maximum value is site was used to obtain the data. There are two groups of brain MR
used to estimate prediction confidence) exceeds a certain thresh- pictures in the databases (normal, glioma tumour). Each dataset is
separated into a training set (which accounts for 70% of the whole
old θk . Assume that q throughout all layers is constant, which lets
dataset) and a test set (which accounts for 30% of the total dataset)
us to calculate the chance that a sample will depart at classifier k
(30 percent of the total dataset).
as:
Fig. 4 illustrates the filtered MR image of the brain tumour. An-
alyzing an image with an appropriate tumour of the brain and
qk = z (1 − q)k−1 q (34)
identifying the most affected region is the essential one. Thus fil-
Where, z is a normalizing constant that ensures that k p (qk ) tering the image using Histogram normalization which reduces the
= 1. Solve this constraint for q and determine the thresholds θk on bias from the image and bilateral filter removes the noise of im-
a validation set in such a way that approximately validation sam- ages from the tumour affected MR image.
ples exit at the k-th classifier. This can lead to improved quality Fig. 5 depicts the brain Segmentation of the tumour into sep-
of life and better clinical outcomes for both high and low-risk pa- arate compartments: C1: (green) enhancing region; C2: (blue)
tients. necrotic region; C3: T1 abnormality (hypo-intensity region on T1)
excluding enhancing and necrotic regions; C4: FLAIR abnormality
4. Experimentation and result discussion (hyper-intensity region on FLAIR) excluding T1 abnormality; C5:
FLAIR abnormality (hyper-intensity region on FLAIR) excluding T1
The results and discussions of the suggested DL approach for abnormality. The manually marked compartments are shown in
glioblastoma brain tumour survival prediction are presented in this Figure (a). Our segmentation technique categorized the compart-
section. Based on the parameters mentioned as follows, the re- ments in Figure (b).
sults are estimated, i.e., the parameters are Specificity, Sensitivity, Fig. 6, shows the durations of survival following the specific
MSE, PSNR, Segmentation time and Prediction Accuracy. The pro- treatments correlated with age. Patients who had greater risk rat-
posed MFCM-RSGWO-FRCNN is compared with the existing FCM ings had a worse prognosis than those who had lower risk scores,
[28] [31], OTSUS [32], Naive Bayes (NB) [33], SVM [6] [34] algo- according to the survival curves. The risk signature performed ef-
rithms, and detailed survey was published [35]. Using mixed meth- fectively in predicting GBM patients’ survival, according to these
ods to classify brain tumors and modelling of reconstruction using findings.
three-dimensional visualization for brain tumor detection and clas- Fig. 7 illustrates Brain tumour specimens of high and low grade
sification had proposed [38–40]. The system is implemented in the are compared. The suggested approach was used to demonstrate

9
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Fig. 3. Sample MRI Brain Tumour Images from the Dataset.

Fig. 4. Filtered Brain Tumour Image.

Fig. 5. Brain Tumour Segmentation.

segregation between high-grade and low-grade cytology speci- 4.3. Comparative analysis with performance parameters
mens. As seen in Figure, there was considerable crossover between
the spectra of high-grade and low-grade specimens (n = 310 spec- The Specificity, Sensitivity, Precision, Recall, PSNR, Mean Square
imens). Error (MSE), Segmentation time and Prediction Accuracy are the

10
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Fig. 6. Survival Probability.

Fig. 7. High-Grade (HG) versus Low-Grade (LG) specimens.

performance metrics is illustrated by varying the dataset sizes from 45.1 s and 62.78 s respectively. For exactly segmenting the tumour
50 to 250 and is represented below. region, the suggested method is well suited which is demonstrated
in the figure.
4.3.1. Segmentation time
The amount of time it takes to perform image segmentation 4.3.2. Sensitivity
for a preprocessed MRI image is known as time segmentation (S t ). Sensitivity ( S n ) is the measure of successful detection of tu-
Between the segmentation process’s ending time (T t ) and starting mour affected region in the provided MR images. Tests with high
time (I t ), it is the difference. It is stated as an equation (35) and is sensitivity are more effective when the result is positive. It is de-
expressed in seconds (s). noted in equation (36).

St = T t − It (35) T ed
Sn = (36)
T ed + T ded
The results of segmentation time are portrayed in Fig. 8 from
the ranges of 50 to 250, the number of MRI images were used, Where, T ed represents the count of tumour that exists and de-
and it describes that the segmentation time of the proposed deep tected images, T ded be the count of tumour does not exist and not
learning approach is moderately smaller when associated with the detected images.
Naïve Bayes algorithm and conventional Multi-Layer Perceptron Sensitivity analysis of the proposed along with the existing al-
(MLP). The proposed technique takes 10.99 s to segment the im- gorithms concerning the total MR images ranging from the dataset
age, whereas the compared techniques result as 73.97 s, 30.3 s, of 50 to 250 is portrayed in Fig. 9. The experimental results state

11
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Fig. 8. Representation of Segmentation Time (s).

Fig. 9. Existing and Proposed Algorithm analysis for Sensitivity. Fig. 10. Specificity Analysis of the Existing Processes with the proposed Method.

2 |AI ∩ BI|
that the proposed system produces an average of 90% for any DiceIndex = (38)
number of data set ranging from 50 to 250. Comparatively, the
| A I | + |B I |
proposed DL approach is much better in the prediction and classi- Where, A ∈ {0, 1} is tumour region extracted from algorithmic
fication of the overall survival rate. predictions and B ∈ {0, 1} is the experts ground truth. The min-
imum value of the dice coefficient is 0 and the maximum is 1;
4.3.3. Specificity More overlap between images indicates a higher value. The graph-
Specificity ( S p ) is the measure to correctly identify the tumour ical representation of the Dice Similarity Index is shown in the
unaffected region from the set of MR images. A higher level of following figure.
specificity values results from greater survival prediction results. It Fig. 11 graph illustrates the accuracy of segmenting glioblas-
is denoted in equation (37). toma from real MR images using the Dice similarity metric. Addi-
tionally, the dataset’s segmentation accuracy is determined using
T ded Dice similarity metrics, which measure the accuracy of segmen-
Sp = (37)
T ded + T dep tation. The dice similarity index of the proposed technique was
Where, T dep is the count of total MR images that the tumour 98.86, 98.54, 97.39, 97.2, and 98.9.
portion does not exist but detected. The graphical representation
4.3.5. Mean Square Error
of the specificity measure is demonstrated in Fig. 10.
Mean Square Error is the measure of the variation among the
The result analysis of the existing and proposed algorithms for
specificity measure states that the average specificity rate is about values with the average of squares of error along with the squared
77%. Varying the datasets have also produced greater results. For alteration between the estimated values and the actual value in
each data size, the proposed values generated are high. Accord- the prediction model of the quantitative information. It is generally
ingly, the proposed DL approach retrieves better detection of the considered as an exceptional error metric for numerical informa-
non-tumour portion from the given images. tion prediction and is represented as

4.3.4. Dice similarity index MSE = ( E I p − A I p )2 (39)


The dice coefficient or dice similarity index indicates how much Where, E represents the estimated value and A be the actual
overlap there is between the two images value of input MR images I p . Loss function reduces the system

12
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

Fig. 13. Existing and Proposed Algorithms’ Peak Signal to Noise Ratio Analysis.
Fig. 11. Dice Similarity Index.

Fig. 14. Survival Prediction Accuracy Analysis of the Proposed and Existing Algo-
Fig. 12. Mean Square Error Analysis of the Proposed and Existing Algorithms1.
rithms.

performance. The graphical representation of the MSE loss function


compensated. The bilateral filter’s PSNR is about 45% higher than
is described in Fig. 12.
the compared FCM, SVM, NB and OTSUS methods.
The observation from Fig. 6 is that for each varied data size
of 50 to 250, the mean square error value is 78% than the com-
pared techniques such as FCM, SVM, NB and OTSUS. This measure 4.3.7. Survival prediction accuracy
maximizes the impact of the segmented region in the set of data The number of correctly categorized MRI images divided by the
through the square process to assess the ability of a method to total MRI images is known as accuracy. This measure is also de-
state these regions. fined as the successful prediction of survival rate from the brain
tumour classification. In percentage (%) that can be expressed and
4.3.6. Peak Signal to Noise Ratio (PSNR) is calculated as follows,
After executing the preprocessing task, the PSNR is well-defined
NIc
as the squared difference between the noisy image and the qual- Acpre = (41)
ity enhanced MRI image. In the decibel value (dB), the PSNR is TIMRI
expressed. Equation (40) represents the PSNR calculation. Where, Acpre is the prediction accuracy, NIc be the number of
2 correctly predicted tumour MR images and TIMRI denotes the entire
p
PSNR = 10log10 (40) number of images availed for processing of the DL method.
( E I p − A I p )2 Survival prediction analysis of the proposed along with the ex-
With relation to MSE, p specifies the potential pixel values isting algorithms concerning the amount of MR images ranging
maximum number for MR images. from the dataset of 50 to 250 is illustrated in Fig. 14. The exper-
Fig. 13 shows the PSNR values for MRI image sizes ranging from imental results state that for each varied dataset, the accuracy of
50 to 250 pixels. When comparing the suggested MFCM-RSGWO- prediction is about 20% more than the compared techniques. That
FRCNN to the novel classification and survival prediction model, is for the data size of 50, the proposed technique produces 72%,
the PSNR value of the proposed value is comparatively higher. This whereas the existing techniques produce 70%, 65%, 62% and 58%
is due to the bilateral filtering application with additional normal- respectively for FCM, SVM, NB and OTSUS methods. Similarly, for
ization function in brain MR images, the unique noise features are all the other sets of data, the proposed method produces an av-

13
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

erage of about 95% accuracy in the overall survival prediction of References


brain tumour.
[1] Hein Tun Zaw, Noppadol Maneerat, Khin Yadanar Win, Brain tumour detection
based on Naïve Bayes classification, in: 2019 5th International Conference on
5. Conclusion and future scope Engineering, Applied Sciences and Technology (ICEAST), IEEE, 2019, pp. 1–4.
[2] Jie Fu, Kamal Singhrao, Xinran Zhong, Yu Gao, Sharon X. Qi, Yingli Yang, Dan
Ruan, John H. Lewis, An automatic deep learning–based workflow for glioblas-
Glioblastoma is a dreadful and common type of brain tumour.
toma survival prediction using preoperative multimodal MR images: a feasibil-
A GBM patients’ precise pre-operative prognosis is strongly needed ity study, Adv. Radiat. Oncol. 6 (5) (2021) 100746.
for tailored treatment. As a result, in improving glioma patient [3] Javaria Amin, Muhammad Sharif, Nadia Gul, Mudassar Raza, Muhammad Al-
survival rates, therapy planning is critical. To diagnose activities mas Anjum, Muhammad Wasif Nisar, Syed Ahmad Chan Bukhari, Brain tumour
and problems in the human body before undergoing lengthy op- detection by using stacked autoencoders in deep learning, J. Med. Syst. 44 (2)
(2020) 1–12.
erations, medical practitioners and researchers can use imaging [4] Asgeir Store Jakola, Lisa Millgård Sagberg, Sasha Gulati, Ole Solheim, Advance-
techniques. For patient diagnosis, the information obtained from ments in predicting outcomes in patients with glioma: a surgical perspective,
an MRI scan is sufficient. By radiation therapy or chemotherapy, Expert Rev. Anticancer Ther. 20 (3) (2020) 167–177.
it is the standard treatment for GBM which is surgical excision. [5] Prathibha Goriparthi, Madhav V. Srinu, M. Narendra, Brain tumour classification
using convolution neural networks, in: Recent Advances in Computer Based
Prognosis varies widely among patients, resulting in an extensive Systems, Processes and Applications, CRC Press, 2020, pp. 135–143.
range of overall survival (OS) times, because of the inherent het- [6] Yalanati Ayyappa, Ashok Bekkanti, Azmira Krishna, P. Neelakanteswara, Cmak
erogeneity of GBM. The conventional techniques experimented to Zeelan Basha, Enhanced and effective computerized multi layered perceptron
date produces many challenges such as overhead, data overfitting based back propagation brain tumour detection with Gaussian filtering, in:
2020 Second International Conference on Inventive Research in Computing Ap-
problem, human level errors and so forth. To overcome the above-
plications (ICIRCA), IEEE, 2020, pp. 58–62.
mentioned issues, a dynamic DL approach for GB brain tumour [7] Benjamin A. Hoff, Benjamin Lemasson, Thomas L. Chenevert, Gary D. Luker,
survival prediction was proposed. The first step was preprocessing. Christina I. Tsien, Ghoncheh Amouzandeh, Timothy D. Johnson, Brian D. Ross,
Intensity normalization by histogram normalization and de-noising Parametric response mapping of FLAIR MRI provides an early indication of pro-
gression risk in glioblastoma, Acad. Radiol. (2020).
utilizing bilateral filter has been done to enhance the images, the
[8] Muhammad Sharif, Javaria Amin, Mudassar Raza, Muhammad Almas Anjum,
data impurities were removed. The noises are Gaussian noise and Humaira Afzal, Shafqat Ali Shad, Brain tumour detection based on extreme
salt pepper noise were also removed. Secondly, segmentation was learning, Neural Comput. Appl. (2020) 1–13.
carried out using the MFCM clustering algorithm for radiomic fea- [9] Gopal S. Tandel, Antonella Balestrieri, Tanay Jujaray, Narender N. Khanna, Luca
Saba, Jasjit S. Suri, Multiclass magnetic resonance imaging brain tumour clas-
ture segmentation. Next, the most significant and informative fea-
sification using artificial intelligence paradigm, Comput. Biol. Med. 122 (2020)
tures from the extracted features were selected utilizing Rough Set 103804.
Theory-based Grey Wolf Optimization. Then the total survival pre- [10] Ahmad Chaddad, Siham Sabri, Tamim Niazi, Bassam Abdulkarim, Prediction of
diction classification is applied on feature selected images utilizing survival with multi-scale radiomic analysis in glioblastoma patients, Med. Biol.
Eng. Comput. 56 (12) (2018) 2287–2300.
FR-CNN.
[11] Cheng Fan, Fu Xiao, Yang Zhao, A short-term building cooling load prediction
Experiments are conducted by comparing the MFCM-RSGWO- method using deep learning algorithms, Appl. Energy 195 (2017) 222–233.
FRCNN method with the well-known FCM, OTSUS, NB, SVM al- [12] N. Varuna Shree, T.N.R. Kumar, Identification and classification of brain tumour
gorithms. The method was evaluated based on metrics such as MRI images with feature extraction using DWT and probabilistic neural net-
work, Brain inform. 5 (1) (2018) 23–30.
Specificity, Sensitivity, PSNR, Mean Square Error (MSE), Segmenta-
[13] Li Sun, Songtao Zhang, Lin Luo, Tumour segmentation and survival prediction
tion time and Prediction Accuracy. As well as the advantage of less in glioma with deep learning, in: International MICCAI Brainlesion Workshop,
convergence, the MFCM-RSGWO-FRCNN has the following proper- Springer, Cham, 2018, pp. 83–93.
ties. [14] Hidir Selcuk Nogay, Hojjat Adeli, Machine learning (ML) for the diagnosis of
autism spectrum disorder (ASD) using brain imaging, Rev. Neurosci. 31 (8)
(2020) 825–841.
• It took a lesser segmentation time of about 10 sec for the en- [15] Gökay Karayegen, Mehmet Feyzi Aksahin, Brain tumour prediction on MR im-
tire system processing and is capable of classification-based ages with semantic segmentation by using deep learning network and 3D
prediction of survival rate. imaging of tumour region, Biomed. Signal Process. Control 66 (2021) 102458.
• It is much more effective with the reduced error rate of 2.3% [16] Mohamed A. Naser, M. Jamal Deen, Brain tumour segmentation and grading of
lower-grade glioma using deep learning in MRI images, Comput. Biol. Med. 121
MSE and that of other techniques produces 4.6% and 5% re- (2020) 103758.
spectively. [17] Nikhilanand Arya, Sriparna Saha, Multi-modal classification for human breast
• The system produces greater accuracy in classification and pre- cancer prognosis prediction: proposal of deep-learning based stacked ensemble
diction of about 95% and a reduced noise ratio for the entire model, IEEE/ACM Trans. Comput. Biol. Bioinform. (2020).
[18] Zhenyu Tang, Yuyun Xu, Lei Jin, Abudumijiti Aibaidula, Junfeng Lu, Zhicheng
set of images for any varied data sizes. Jiao, Jinsong Wu, Han Zhang, Dinggang Shen, Deep learning of imaging pheno-
type and genotype for predicting overall survival time of glioblastoma patients,
In comparison with the methods such as FCM, OTSUS, NB, SVM, IEEE Trans. Med. Imaging 39 (6) (2020) 2100–2109.
the efficiency of the MFCM-RSGWO-FRCNN can predict appropri- [19] A. Rehman, M.A. Khan, T. Saba, Z. Mehmood, U. Tariq, N. Ayesha, Microscopic
brain tumour detection and classification using 3D CNN and feature selection
ate images in the data source. The MFCM-RSGWO-FRCNN was also
architecture, Microsc. Res. Tech. 84 (1) (2021) 133–149.
found more robust than the FCM, which was proposed to segment [20] Yannick Suter, Urspeter Knecht, Mariana Alão, Waldo Valenzuela, Ekkehard
based on clustering. While comparing the performance with the Hewer, Philippe Schucht, Roland Wiest, Mauricio Reyes, Radiomics for glioblas-
OTSUS, NB and SVM produce better quality outcomes. The pro- toma survival analysis in pre-operative MRI: exploring feature robustness, class
boundaries, and machine learning techniques, Cancer Imaging 20 (1) (2020)
posed algorithm could be extended in several ways. The method
1–13.
may first do a detailed kind, time-based tumour presence, loca- [21] Navodini Wijethilake, Mobarakol Islam, Hongliang Ren, Radiogenomics model
tion, and size analysis. As a result, several difficulties have arisen, for overall survival prediction of glioblastoma, Med. Biol. Eng. Comput. 58
including the usage of different validity indices, dissimilarity met- (2020) 1767–1777.
[22] Parita Sanghani, Beng Ti Ang, Nicolas Kon Kam King, Hongliang Ren, Over-
rics, and comparisons with other hybrid algorithms for reducing
all survival prediction in glioblastoma multiforme patients from volumetric,
segmentation time can be completed later on. shape and texture features using machine learning, Surg. Oncol. 27 (4) (2018)
709–714.
[23] Navodini Wijethilake, Dulani Meedeniya, Charith Chitraranjan, Indika Perera,
Declaration of competing interest
Survival prediction and risk estimation of glioma patients using mRNA ex-
pressions, in: 2020 IEEE 20th International Conference on Bioinformatics and
There is no conflict of interest Bioengineering (BIBE), IEEE, 2020, pp. 35–42.

14
D.S. Wankhede and R. Selvarani Neuroscience Informatics 2 (2022) 100062

[24] Ujjwal Baid, Swapnil U. Rane, Sanjay Talbar, Sudeep Gupta, Meenakshi H. volutional neural network, in: Data Visualization and Knowledge Engineering,
Thakur, Aliasgar Moiyadi, Abhishek Mahajan, Overall survival prediction in Springer, Cham, 2020, pp. 171–194.
glioblastoma with radiomic features using machine learning, Front. Comput. [33] Gurkarandesh Kaur, Ashish Oberoi, Novel approach for brain tumour detection
Neurosci. 14 (2020) 61. based on Naïve Bayes classification, in: Data Management, Analytics and Inno-
[25] Vikas Kumar Anand, Sanjeev Grampurohit, Pranav Aurangabadkar, Avinash Kori, vation, Springer, Singapore, 2020, pp. 451–462.
Mahendra Khened, Raghavendra S. Bhat, Ganapathy Krishnamurthi, Brain tu- [34] S. Deepak, P.M. Ameer, Automated categorization of brain tumour from MRI
mour segmentation and survival prediction using automatic hard mining in using CNN features and SVM, J. Ambient Intell. Humaniz. Comput. 12 (8) (2021)
3D CNN architecture, in: International MICCAI Brainlesion Workshop, Springer, 8357–8369.
Cham, 2020, pp. 310–319. [35] Disha Wankhede, Selvarani Rangasamy, Review on deep learning approach for
[26] Marcin Kociołek, Michał Strzelecki, Rafał Obuchowicz, Does image normaliza- brain tumor glioma analysis, in: International Conference on Convergence of
tion and intensity resolution impact texture classification, Comput. Med. Imag- Smart Technologies IC2ST-2021, https://doi.org/10.17762/itii.v9i1.144.
ing Graph. 81 (2020) 101716. [36] S.L. Bangare, Classification of optimal brain tissue using dynamic region grow-
[27] Qiwei Xing, Chunyi Chen, Zhihua Li, Progressive path tracing with bilateral- ing and fuzzy min-max neural network in brain magnetic resonance images,
filtering-based denoising, Multimed. Tools Appl. 80 (1) (2021) 1529–1544. Neurosci. Inform. (ISSN 2772-5286) 2 (3) (September 2022) 100019, https://
[28] Jiaqing Miao, Xiaobing Zhou, Ting-Zhu Huang, Local segmentation of images doi.org/10.1016/j.neuri.2021.100019.
using an improved fuzzy C-means clustering algorithm based on self-adaptive [37] N. Shelke, S. Chaudhury, S. Chakrabarti, S.L. Bangare, et al., An efficient way
dictionary learning, Appl. Soft Comput. 91 (2020) 106200. of text-based emotion analysis from social media using LRA-DNN, Neurosci.
[29] B. Sathiyabhama, S. Udhaya Kumar, J. Jayanthi, T. Sathiya, A.K. Ilavarasi, V. Yu- Inform. (ISSN 2772-5286) 2 (3) (September 2022) 100048, https://doi.org/10.
varajan, Konga Gopikrishna, A novel feature selection framework based on grey 1016/j.neuri.2022.100048.
wolf optimizer for mammogram image analysis, Neural Comput. Appl. (2021) [38] S.L. Bangare, G. Pradeepini, S.T. Patil, Brain tumor classification using mixed
1–20. method approach, in: 2017 International Conference on Information Commu-
[30] Riccardo Rosati, Luca Romeo, Sonia Silvestri, Fabio Marcheggiani, Luca Tiano, nication and Embedded Systems (ICICES), 2017.
Emanuele Frontoni, Faster R-CNN approach for detection and quantification of [39] S.L. Bangare, Brain tumor detection using machine learning approach, Des. Eng.
DNA damage in comet assay images, Comput. Biol. Med. 123 (2020) 103912. (Scopus Index-Q4) (ISSN 0011-9342) 7 (2021) 7557–7566.
[31] A. Biswas, M.S. Islam, ANN-based brain tumour classification: performance [40] S.L. Bangare, G. Pradeepini, S.T. Patil, Implementation for brain tumor detection
analysis using K-means and FCM clustering with various training functions, in: and three dimensional visualization model development for reconstruction, J.
Explainable Artificial Intelligence for Smart Cities, CRC Press, pp. 83–102. Eng. Appl. Sci. (ISSN 1819-6608) 13 (2) (2018) 467–473.
[32] Surbhi Vijh, Shilpi Sharma, Prashant Gaurav, Brain tumour segmentation us-
ing OTSU embedded adaptive particle swarm optimization method and con-

15

You might also like