International Journal of Intelligent Networks 5 (2024) 241–254

Contents lists available at ScienceDirect

International Journal of Intelligent Networks

journal homepage: www.keaipublishing.com/en/journals/
international-journal-of-intelligent-networks

Revolutionizing prostate cancer diagnosis: Unleashing the potential of an

optimized deep belief network for accurate Gleason grading in
histological images
S. Angel Latha Mary a, *, S. Siva Subramanian b, G. Priyanka c, T. Vijayakumar d,
Suganthi Alagumalai e
a
IT Department, SNS College of Technology, Coimbatore, India
b
Department of of BME, Rathinam Technical Campus, Coimbatore, India
c
Department of Computer Science and Business Systems, Sri Eshwar College of Engineering, Coimbatore, India
d
Department of AIDS, Dr.Mahalingam College of Engineering and Technology, Pollachi, Coimbatore, India
e
Cyber Operations, 3223 Hanover St, Palo Alto, CA, 94304, USA

A R T I C L E I N F O A B S T R A C T

Keywords: PC (Prostate Cancer) is the second highest cause of death due to cancer in men globally. Proper detection and
Prostate cancer treatment are critical for halting or controlling the growth and spread of cancer cells within the human organism.
Gleason grading However, evaluating these sorts of images is difficult and time-consuming, requiring histopathological image
Histological images
recognition as the most reliable method for treating PC because of its distinct visual characteristics. Risk eval-
Elephant herding optimization
Hyper-parameter and Gaussian kernels
uation and treatment planning rely heavily on histological image-based Gleason grading of prostate tumors. This
work introduces an innovative approach to histological image analysis for prostate cancer diagnosis and Gleason
grading. The Elephant Herding Optimization-based Hyper-parameter Convolutional Deep Belief Network (CDBN-
EHO) is presented alongside a grading network head-optimized deep belief network technique for multi-task
prediction. Leveraging an effective Bayesian inference method, fully linked Conditional Random Field (CRF)
techniques are utilized for segmentation, with pairwise boundary capacities determined by a linear mixture of
Gaussian kernels. The multi-task approach aims to enhance performance by incorporating contextual informa-
tion, leading to breakthrough results in the identification of epithelial cells and the grading of Gleason scores.
The objective of this study is to demonstrate the effectiveness of the optimized deep belief network technique in
improving diagnostic accuracy and efficiency for prostate cancer diagnosis and Gleason grading in histological
images.

1. Introduction surgery had a higher chance of surviving the disease (91 %). These
findings demonstrate that despite the fact that therapy has the potential
In India, PC ranks among the top 10 most common malignancies. to save a life or increase the number of years a person can live, it is more
Men over the age of 65 are most likely to be affected by this condition. important than ever to raise awareness about the disease and take steps
However, there has been an increase in the number of cases of cancer to prevent it. Estimates of the number of people diagnosed with PC in
reported in younger men living in urban areas who are between the ages India for the years 2010 and 2015 were respectively 26,120 and 28,079
of 35 and 44 and between 55 and 64 in recent years. Some of the most [1]. The number of new instances of this disease is expected to more
significant factors that have been found to increase the risk of prostate than quadruple by the year 2020, according to the data that was gath-
cancer are becoming older, being overweight, eating an unhealthy diet, ered from cancer projections. Incidence rates of this type of cancer are
and having genetic changes. Cancer has a 64 % cure rate after 5 years in consistently and rapidly increasing.
patients in India. According to the findings of a study that was carried Artificial Intelligence (AI) methods are computer algorithms that
out in Mumbai, patients who received PC treatment that also included learn patterns from previously collected data in order to qualify

* Corresponding author.
E-mail addresses: [email protected] (S. Angel Latha Mary), [email protected] (S. Siva Subramanian), [email protected] (G. Priyanka),
[email protected] (T. Vijayakumar), [email protected] (S. Alagumalai).

https://doi.org/10.1016/j.ijin.2024.05.004
Received 2 September 2023; Received in revised form 18 April 2024; Accepted 4 May 2024
Available online 18 May 2024
2666-6030/© 2024 The Authors. Published by Elsevier B.V. on behalf of KeAi Communications Co., Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

predictions in novel data that has not been seen. Earlier iterations of AI illustrated in Fig. 1 and the key contribution of this work.
made use of what are now commonly referred to as ‘conventional’ ma-
chine learning (ML) methodologies. These procedures were frequently • To introduce a novel CDBN-EHO and a specialized deep belief
carried out in two stages. To begin, domain experts (humans who are network technique tailored for multi-task prediction, particularly in
specialists in the subject area) painstakingly created features to excerpt the context of segmentation tasks.
numerical characteristics from the data that were particular to the task • This work also presents an advanced Bayesian inference method
at hand. For instance, tumor size or form was one of the variables that applied to fully linked CRF techniques for segmentation, where
was extracted. Second, these manually produced features were input pairwise boundary capacities are determined using a linear mixture
into automated detection techniques in order to discover which aspects of Gaussian kernels.
were helpful and how to integrate them in order to achieve the highest • By adopting a multi-task learning approach, the proposed techniques
level of accuracy possible when classifying data into categories (for aim to enhance performance by leveraging shared contextual infor-
example, benign nodules versus malignant tumors). These AI methods mation across tasks. Notably, these methods have achieved
can be trained to make predictions based on data they have never seen remarkable breakthroughs in the simultaneous identification of
before, and then utilized to make predictions in novel contexts. Deep epithelial cells and grading of Gleason scores.
learning (DL) approaches were made possible by recent advances in the
computational capability of graphics processing units, commonly Overall, the combination of these techniques represents a significant
referred to as GPUs. Because deep learning techniques eliminate the advancement in the field of machine learning, particularly in the do-
requirement for hand-crafted features, they can function in a completely mains of multi-task prediction, segmentation, and grading. The
computerized manner to not only detect the characteristics but also use demonstrated breakthrough results underscore the effectiveness and
them for the task that is wanted further down the line. DL techniques potential applicability of the proposed methodologies in various real-
have brought about a sea change in the field of AI as a result of their world scenarios.
exceptional performance, which frequently exceeds that of humans, The remaining of the paper is structured as follows: the related work
notably in activities linked to image processing. PC using DL is discussed in section 2. The proposed methodology is
The field of medical imaging, especially detection using automated explained in section 2 and the experimental results and analysis is
assessment, which is the incorporation of imaging function architecture explained in section 4. The final conclusion and future work are given in
based on machine learning, has shown that it can help radiologists make section 5.
correct diagnoses, which cuts down on diagnostic time and also
screening costs. The USG (Ultra Sono-Grapy) technique is a sort of im-
aging technology used in the first detection step and is based on real-
time images. This method, however, cannot yield the best outcomes
due to the absence of contrast between cancer and benign tissue [2].
Despite its lack of real-time imaging capability, mpMRI (multi-
parametric magnetic resonance imaging) tends to be more precise than
ultrasound due to superior tissue contrast. However, in order to use it,
extensive training is necessary [3].
The areas of the skull and neck, cervical, and pelvic cavity can all be
treated with brachytherapy, which is a form of radiation therapy. With
this method, the radioactive beam is aimed directly at the area of the
body that has been damaged. There are two different approaches to
brachytherapy: permanent installation, which is also referred to as seed
placement, and removal of radioactive material [4]. It has been found
that brachytherapy, which involves the use of seeds, is more real in
treating prostate cancer than radical prostatectomy. In a procedure
called brachytherapy, which is used to operate on the prostate, pallets,
which are little seeds, are positioned on the prostate cell. These pallets
are quite small but contain radioactive materials. This highly radioactive
substance is toxic to PC cells and destroys them. It is possible that this
will cause destruction or even death to the strong cells that are located
close to the prostate as a side effect.
The results of applying DL strategies to various applications, such as
object identification, segmentation, and classification, have been
encouraging. These methods involve the use of convolution layers, each
of which is able to excerpt a unique set of features from the input pho-
tographs by using local low-level features as a starting point and
working their way up to global high-level features. At the boundary of
the coevolutionary neural layers, a fully linked layer transforms
complicated features in terms of signal likelihood [5,6]. The batch
normalization layer, which adjusts the provided input layer with zero
means, and the unit variant with a dropout layer, which appears to be
one of them of the normalization methods that disregard the nodes that
are chosen in a randomly directed fashion, have each been given as
various methods with which to enhance the output based on a DL-based
technique. With these two layers, the output has been enhanced. How-
ever, in order to attain convincing efficiency, the appropriate combi-
nations and layer topologies, along with appropriate fine-tuning of the
hyper-parameters, are required. The following architecture, which is Fig. 1. Architecture diagram of Proposed Methodology for PC detection.

242
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

2. Related work Table 1

The research gap of existing methods with advantages and disadvantages.
The LSTM (long short-term memory) and Residual Net (ResNet - 101) Author Methods Advantages Disadvantages
were used by Iqbal et al. [7], and both were trained without the use of
Iqbal et al. LSTM- Due to the contrast It cannot detect
any manually-created features and were then fine-tuned. Results from [7], ResNet- between malignant and prostate cancers in
non-DL classifiers like the Gaussian Kernel, SVM (support vector ma- 101 benign tissue, these other areas.
chine), kernel naive Bayes, KNN - Cosine (k-nearest neighbor-Cosine), procedures can yield ideal
decision tree (DT), and RUSBoost tree were compared to features that results.
Li et al. [8], Deep Reproducibility is This method had
were created by humans. Non-DL techniques with GLCM features uti- Neural restricted and can be obtained good results,
lizing KNN-Cosine achieved the best results, with 98.01 % sensitivity, Network applied fast in medical but it required precise
99.25 % specificity, 98.99 % PPV, 99.11 % NPV, 99.07 % accuracy, and systems. localization of small
0.998 AUC. With DL method ResNet - 101, received (100 %) Accuracy image portions in order
to extract features.
and AUC (1) for Kernel Naive Bayes, RUSBoost Tree, and SVM Gaussian,
Karimi et al. CNN This approach could have However, developing
whereas DL method LSTM produces performance with specificity (100 [9], reduced unpredictability this approach is not a
%), sensitivity (98.33 %), NPV (99.26 %), PPV (100 %), MCC (0.9879), in the production of large straightforward task
accuracy (99.48 %), and AUC (0.9999). numbers of slides and because significant
Researchers Li et al. For PC pathological images, Furthermore pre- increased their efficiency. volumes of labeled
training data are
sented a DL-based method for automatically determining the Gleason necessary.
grading and segmenting the Gleason outline region. To ensure precise Duran-Lopez CNN • By reducing the amount With a CNN-based
Gleason grading, Furthermore present an architecture for segmenting et al. [10], of data sampled during setup, the
the Gleason pattern region that combines atrous spatial pyramid pooling the generation and computational cost of
classification phases, anticipating an input
with multiscale typical convolution. In addition, the prediction un-
this technique is able to image is significantly
dergoes a post-processing approach employing conditional random speed up the learning higher, resulting in a
fields. For the area covered by the Gleason pattern, the average%age of and inference processes. longer processing time.
correct intersections over unions is 77.29 %, and the overall pixel ac- • The use of a digital
curacy is 89.51 %. In addition, the automatic Gleason grading yielded pathology image to
enhance the CNN’s
results that were on par with those obtained by human pathologists.
adaptation for
Cohen’s quadratic kappa revealed an average inter-annotator agreement classification tasks.
among the technique and the pathologists of 0.77. The proposed Kott et al. CNN The sensitivity and However, their effect is
research demonstrates that merging various DNN (Deep Neural [11], specificity of this diminished due to
approach were both difficulties in
Network) architectures is an effective strategy for producing more de-
improved when applied to reproducing and
tached and predictable Gleason grading of PC. cancers. validating their
Karimi et al. [9] proposed using a DL-based method of classification findings.
and data augmentation methods and accurate grading of PCa in histo- Busby et al. SNN Higher representation Results were greatly
pathology images. The proposed method integrates the outputs of three [12], capacity describes this exaggerated, and their
approach. Statistically, the significance was
independent CNNs (Convolutional Neural Networks) that are trained on
effect of missed.
progressively smaller patches. A logistic regression technique, trained underperformance on
independently of the CNNs, is then used to combine the predictions from other data was substantial.
the three CNNs. Then propose new data augmentation strategies and Ahmad et al. CNN This technique requires a • However, present
[13], lot of computer power and studies cannot
conduct empirical research into their influence on the classification
a powerful graphics answer if a PC
accuracy to better train This algorithms. The suggested technique processing unit. symptom arises on
correctly identifies 92 % of malignant patches as cancerous rather than any portion of the
benign, and correctly 86 % of low-grade cancers. body.
An innovative DL-based CADS was presented by Duran-Lopez et al., • This approach
delivers adequate
[10]. After being patch-sampled and pre-processed with several filters,
accuracy but
including a revolutionary patch-scoring method that gets rid of unim- necessitates clean,
portant parts of the tissue, this system can interpret whole-slide histol- asymmetric, and
ogy images. The consequence that using a stain-normalization technique sizable training
datasets.
on the patches to lessen color variation between scanners is also
Swiderska- GAN • This strategy has a Continuously providing
analyzed. Utilizing 3-fold cross-validation for training, the network Chadaj greater chance of varied image data is
achieves 0.999 AUC, 0.999 F1 score, and 99.98 % accuracy on a new test et al. [14], success. required for testing
set. The average processing time for generating a heatmap of an entire • The data produced by purposes.
slide is 15 s. For a binary classification job between normal and GANs resembles the
original data quite
cancerous prostate whole-slide images, This proprietary network beats
closely and can be
prevailing advanced research in terms of computational complexity at readily interpreted into
the patch level. variants.
Using a deep residual CNN, Kott et al. [11] were able to classify each Kumar et al. VGGnet This technique allows for High cost in carrying
patch on two different levels: (1) broadly (benign vs. malignant) and (2) [15], precise remote prediction out the modifications;
at a low cost and with yet, the significance
specifically (benign vs. Gleason 3 vs. 4 vs. 5). Furthermore used lightning-fast and relevance were not
five-sample cross-validation to assess the quality of This techniques. comprehension. completely examined
Hypothesis testing comparing actual technique performance with that because of a lack of
expected by chance was conducted using randomization tests. comprehensive.
Chatrian DNN Highly substantial • But pathologic
Fine-grained classification of image patches as benign or malignant
et al. [16], changes in the forecasting complete response
showed 91.5 % accuracy (p 0.001) using the technique (0.90 specificity, of overall survival when rates were not
0.93 sensitivity, and 0.95 average precision). The requirement for (continued on next page)
external validation and the relatively modest size of the sample are two

243
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

Table 1 (continued )
Author Methods Advantages Disadvantages

using this approach to significantly

determine the prostate different.
size. • Poor contrast
between hard and
soft tissues.

limitations. Heterotopic choriocarcinoma was successfully diagnosed

and graded using a DL-based computer vision algorithm in this study.
According to Busby et al. Shortages in the available workforce and
inconsistencies in histopathological evaluation. These techniques,
which include histopathological information into complex neural net-
works, perform exceptionally well at identifying, grading, and predict-
ing outcomes for PC. In spite of the fact that fully autonomous PC
diagnosis is still years away, new research indicates that AI is already
being used as a preliminary screening tool, a companion in the form of a
real-time interactive screen during histological analysis, and an addi-
Fig. 3. Outline of the CDBN-EHO technique proposed architecture.
tional read system to identify false negative diagnoses. The persistence
of this work is to discuss the current state of AI in PC histopathology as
well as its potential future applications. (whole-slide image color standardizer) technique. Furthermore, found
Histopathologic images of lymph node sections were used by Ahmad an AUC of 0.92 and 0.83 for the two data sets that were treated sepa-
et al. [13], to present a CNN-based technique for the categorization and rately. The AUC increases from 0.91 to 0.88 when rescanning, and from
diagnosis of metastatic cancer. The process of diagnosing cancer from 0.97 to 0.98 when normalizing for formatting. This system may one day
histopathologic images is laborious and time-consuming for pathologists be implemented to mechanically pre-screen prostate biopsies, relieving
because a broad tissue area must be investigated, and even a single pathologists of some of their duties.
microscopic metastasis can be missed. Furthermore, took the required According to Kumar et al. Using differential privacy and safe multi-
precautions against overfitting and improved the findings by doing the party computation, a robust framework was proposed for cancer diag-
appropriate pre-processing and data augmentation activities. The nostic picture classification. Instead of doing the entire operation in the
method achieves excellent accuracy in cancer diagnosis through the use cloud, one can split the layers apart into two distinct modules: a single
of low-dimensional representations and automated, specific feature for feature extraction using the VGGNet module on the user side, and the
extraction and classification. The experimental results for the medical other modules for private prediction in the cloud. Two data sets made up
image categorization and detection challenge demonstrate promising of histological images of canine mammary tumors and human breast
performance, with an accuracy rate of 0.94. cancer are used to verify the framework’s efficiency. When applied to
Swiderska-Chadaj et al. [14] analyzed the impact of imaging devices the suggested technique, differential privacy preservation makes it
and cycle-GAN-based standardization on the efficacy of algorithms and secure and able to protect private information from prying eyes without
evaluated several DL techniques for discovering PC in whole-slide im- sacrificing performance. The suggested approach efficiently strikes a
ages. Among the many networks tested in this work are U-Net, Dense- balance between privacy and technique performance, as demonstrated
Net, and EfficientNet. As a pre-processing phase, furthermore by a battery of experiments.
investigated the cycle-GAN-based normalizing technique and the WSICS In a recent study, Chatrian et al. [16] showed how variational

Fig. 2. Original, benchmark, and problem-defining sample images.

244
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

autoencoders and generative adversarial networks may be used together on more complex cases and critical decision points.
to generate realistic histological images that can be used to train se-
mantic segmentation techniques. Then examine if these techniques can
be employed to separate out groups of prostate glands that share a 3.1. Dataset description
common molecular signature. If this progress continues, it will lead to
the identification of previously unknown histology-based subgroups of Grading histological images for Gleason scores is a challenging and
diseases. It shows that the expression of a clinical indicator in prostate time-consuming process that demands expertise and experience from
glands can be identified using only morphological parameters extracted pathologists. However, the development of automated methods offers a
from H&E images. promising solution. These advanced techniques have the potential to
Summary: Various studies have tackled prostate cancer diagnosis accurately analyze histological images, providing valuable assistance to
and Gleason grading using a range of techniques, including non-deep pathologists in their diagnostic workflow. By automating this task,
learning (DL) methods like Gaussian Kernel, SVM, and DL methods workload can be reduced, and consistency in grading can be improved,
such as ResNet-101, LSTM, and CNNs. While non-DL techniques like ultimately benefiting both pathologists and patients. IRB number
GLCM features achieved high sensitivity and specificity, DL methods like Pro00029960 authorizes the use of 513 images from the Pathology
LSTM demonstrated excellent performance with high accuracy and AUC Department archives at Cedars-Sinai Medical Center as the basis for the
shown in Table 1. Innovations in DL-based approaches, such as inte- proposed dataset. The 513 images are a mosaic made up of two different
grating CNNs and employing novel architectures, have led to significant kinds of tiles. Twenty-four photos (Set A) from these individuals have
advancements in automated Gleason grading and histological image been classified as having either stroma (ST), benign or normal glands
interpretation. Despite the progress, challenges remain, including the (BN, scored as GG2 or lower), LG cancer (low-grade), or HG cancer
need for further validation and addressing computational complexity. (high-grade) [17]. The remaining 289 images come from 20 patients and
Nonetheless, AI applications show promise as screening tools and depict various types of dense high-grade tumors, including both cribri-
companions to pathologists, indicating a potential future for AI-assisted form and non-cribriform glands with Gleason grades 5 (GG5) and 4
prostate cancer diagnosis. Traditional deep learning methods like CNNs, (GG4). Set B of these photos (included in Ref. [18]) consists entirely of
LSTM, and GAN have played crucial roles in image analysis tasks like stromal components including nerve tissue and blood arteries. Set A
classification and segmentation. However, they may fall short in fully slides were scanned using a high-resolution whole-slide scanner
leveraging contextual cues or handling multi-task prediction challenges, SCN400F (Leica Biosystems, Buffalo Grove, IL), and Set B slides were
particularly in histological image analysis. Also, their computational scanned using the Aperio scanning system (Aperio ePathology Solutions,
complexity may hinder their efficacy, especially in addressing the in- Vista, CA). Both systems used a 20× scanning objective. The final
tricacies of multi-task prediction, notably in histological image analysis. product was a color RGB image with an 8-bit intensity depth per color
In contrast, the proposed method introduces innovative optimization network and a pixel extent of 0.5 m 0.5 m. Whole Slide Images (WSIs)
techniques, such as EHO and hyper-parameter optimization, specifically were partitioned into 1200 × 1200-pixel tiles after representative tiles
tailored for multi-task prediction. Moreover, optimizing the DL archi- were manually selected by a pathologist. An experienced research
tecture for grading tasks reflects a targeted strategy for addressing the pathologist used a custom-built graphical user interface to manually
intricacies of Gleason grading in histological images. Through an annotate the data included in each tile. Three samples from the dataset
emphasis on multi-task learning and task-specific optimizations, the utilized in this work are displayed in Fig. 2. The pathologists checked
proposed method aims to elevate the performance of deep learning each other’s work and agreed on how to fix any discrepancies they found
models in histological image analysis, potentially enhancing diagnostic in the annotated image tiles. To accommodate for stain variation, all
accuracy and aiding clinical decision-making in fields like prostate tiles had been normalized before further analysis. Images on the tiles
cancer diagnosis. were flipped, mirrored, and rotated before being transmitted into the
network as a form of data augmentation. Both [17,18] make use of these
3. Proposed work same datasets from prior research. Please see the Supplementary Infor-
mation for a detailed explanation of the Gleason grading system and the
As a means of addressing the issue raised in the introduction, this rationale for the choice to divide tissues into four groups.
part provides an in-depth analysis of the unique CDBN-EHO framework.
Finally, furthermore present the assessment metrics that were used to i. Network architecture
evaluate the proposed technique and compare it to prior attempts.
Automation of tasks such as Gleason grading through advanced deep Fig. 3 depicts the full system along with the individual parts of the
learning techniques can alleviate the burden on pathologists, who suggested technique. For This proposed picture parser, furthermore rely
typically face heavy workloads and time constraints. By streamlining the on ResNet as the main neural network. The picture parser begins by
analysis process, this method allows pathologists to focus their expertise creating feature maps. Two subsystems receive these feature maps as
input. Similar to the CDBN-EHO, here used a two-stage process in the left

Fig. 4. The entire procedure of the proposed CDBN

245
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

fork. First, a RPN (Region Proposal Network) uses the feature maps to computational rate, and a slow learning time. The number of groups that
propose RoIs (regions of interest). To predict the box offset, class, and are contained within the concealed layers is the second free parameter.
binary mask for each RoI, a GNH (Grading Network Head) is employed There is no overarching principle that can be applied when selecting this
in the second step. Furthermore, augment this with a right fork that parameter. A slower learning time and poorer performance are the
provides a score for the presence of epithelial cells in the image. The inevitable outcomes of having more or fewer units, just as with the
ENH (Epithelial Network Head) designates this section. Both the ENH numeral of hidden levels. The numeral of tiny batches becomes the third
and GNH results contribute to the final forecast generated by the unrestricted parameter. Mini batches have the potential to reduce the
network. The prediction is then subjected to a final round of processing amount of computation required by the CDBN network.
in the form of a conditional random field.
3.3. Feature encoding
ii. Objective function of CDBN-EHO
A feature encoding block is provided with the CDBN features in order
The suggested CDBN-EHO technique has two main functions: to to obtain faster performance. This block is responsible for computing
identify epithelial cells and to generate a segmentation mask based on more discriminative representations. As a method for encoding features,
the Gleason score. The CD (Contrastive Divergence) and PCD (Persistent furthermore make use of the algorithm known as the bag of features
CD) are used to learn the weight matrix and corresponding bias vectors [19]. Fig. 4 illustrates the overall structure of the bag of features. First,
of the exposed and concealed nodes. Here, the BP is combined with a the codebook that is a collection of code words, is built. The code words
standard gradient ascent technique to find the best possible settings for represent the CDBN features that have been retrieved. After that, a
the weight matrices. In order to reduce certain error metrics, the opti- histogram that displays the occurrence of the code words contained
mization process takes into account the results of an additional layer inside the image is used to represent the input image.
constructed on top of the DBN following its first greedy training. Logistic A clustering algorithm is utilized in the process of modelling the
units also known as softmax are commonly utilized in this tier. codebook. In order to determine a collection of centroids in the feature
space, each of the retrieved CDBN features is first clustered. Clustering is
accomplished with the help of the k means method [20], which is used in
3.2. Convolutional deep belief network (CDBN) this work. The choice of the total number of code words, often known as
the size of the codebook, is an essential component in the process of
The input images are processed using a whitening technique that building the codebook. Csurka et al. ([21]) have shown that when it
removes the correlation between neighboring intensity levels. The pro- comes to natural image classification, a bigger codebook size yields
cess of whiten projection involves projecting the input image into the superior results. Despite this, Tatiana and colleagues ([22]) have shown
eigenvectors and normalizing them so that all intensity values have a that the dimensions of the codebook do not have a discernible impact on
variance of 1. The image that has been brightened is then specified to the the precision of the medical picture classification. In order to identify the
visible layer of the CDBN, where it is separated at random into tiny codebook size that is optimal for the categorization of neuroblastoma
batches that overlap with each other. Mini batches break up the input histology images, furthermore first examine a variety of codebook sizes
image into manageable pieces before applying the learning algorithm to and then select the one that performs the best. A histogram of code
each one. When the data sample is not representative of the total, these words is generated to serve as a representation of the input image when
techniques reduce the amount of noise in the analysis. By utilizing a the feature encoding block is utilized. Finally, in order to represent the
hierarchical framework, the CDBN technique is able to separate the low- distinct types of neurological tumor histology images, an SVM classifier
level properties from the high-level ones. The Convolutional Restricted is trained to categorize the histograms of the information-encoded unit.
Boltzmann Machines (CDBMs) are generative techniques built from Furthermore, utilize the Kyoto natural image dataset [23] to learn
CRBMs stacked one on top of the other. As can be seen in Fig. 4, each the first layer of the CDBN because it has a high number of low-level
CRBN has a hidden layer (H), a visible layer (V), and a pooling layer. features, including edges. This is necessary because the first layer of
A NV × Nv array of binary units makes up the viewable layer of the the CDBN is responsible for learning the common visual information,
image. There is a total of k × N2H hidden units since the H is composed of such as edges. The second and third layers receive training with a
k groups (also known as “bases”), each of which is an array of binary database consisting of neuroblastic tumors. This is done by randomly
units with the dimensions NH × NH . In the visible layer, there is a dividing the database into three subsets: the first one is used for vali-
convolutional window with the dimensions NW × NW (where NW = dation and contains 211 photos; the second one is used for training and
NV − × NH + 1), which is dedicated to each individual group. A weight contains 623 images; and the 3rd one is used for testing and contains 209
matrix, denoted by the letter W, is used to describe the symmetric re- images. Using the validation set, furthermore determine the optimal
lationships that exist between hidden and visible units. Here might think values for the free parameters, and furthermore utilize the training set
of each weight matrix as a filter if this method wanted to. In addition, and the testing set to do the final system evaluation.
there are k groups of units that make up the pooling layer, and inside
each group, there are NP × NP binary units. A predetermined constant 3.4. Elephant herd optimization (EHO)
factor C is used by the pooling layer to reduce the size of the repre-
sentation of the H. In order to accomplish this goal, the pooling layer In 2015, Wang et al. presented the EHO algorithm to the scientific
picks the values that are the highest in the C × C windows of the con- community [24]. Elephants engage in social behavior and comprise a
cealed layer. The use of maximum pooling not only makes it possible for complicated hierarchy consisting of females and their young. An
the outputs of higher layers to remain invariant to changes of a more elephant group is made up of a number of different clans, each of which
minute magnitude in the input, nonetheless it also lowers the overall is led by a matriarch along with her calves and any other females that are
computing cost. In order to extract features from histological images of connected to her. A family is founded by a female. EHO is concerned
neuroblastoma, furthermore make use of a CDBN with three layers. The with the hypotheses that come after it. The herd of elephants can be
CDBN that has been presented has some free parameters that still need to broken down into subgroups known as clans, with particular elephants
be established. The first factor to consider is the total number of con- constituting each clan. There is a certain number of male elephants that
cealed layers. While a CDBN with less H s requires less time to train, its break away from their herd to live on their own. There is a matriarch
performance suffers as a result of the network’s inability to ignore low- who serves as the leader of each clan. Within the elephant herd, the
level details in the input images. The performance can be improved by matriarchal group guards the most effective strategy. Each and every
increasing the number of H s, but this can lead to overfitting, a high member of the elephant population belongs to one of the j clans. The

246
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

new posture of each elephant is influenced by the matriarch, ci . It is

possible to determine the elephant j in clan ci by utilizing Eq. (1);
( )
xnew , ci,j = xci,j + a × xbest,ci − xci,j × r (1)

where xnew , ci,j denoted the new location and xci,j signified the previous
location for elephant j within the clan ci . The matriarch ci is indicated
by the value xbest,ci , which symbolizes the best elephant. The scaling
factor, r ∈ [0,1], is shown by the expression a ∈ [0,1]. The most valuable
elephant is determined for each clan based on Eq. (2);
xnew , ci, j = β × xcenter, i (2)

In this case, β ∈ [0, 1] denotes the second constraint that directs the
influence of the xcenter,ci,d , which is outlined in the previous sentence.

1 ∑ n
ci

xcenter,ci,d = × xci,j,d (3)

nci j=1

Here 1 ≤ d ≤ D and nci denote the numeral of elephants in clan xci,j,d

represents the dth dimension of individual elephant xci,j,d , the center of
clan xcenter,ci,d can be modified using the equation (eqn. (3)). When trying
to find solutions to optimization problems, the process of separating
could be modeled as a separation operator. In each clan, the elephants
with the lowest value are moved to the position specified in Eq. (4),
which is the next available spot.
xworst,d = xmin + (xmax − xmin + 1) × rand (4)

In this instance, the bottom and upper bands of the exploration space
are denoted by the variables xmin and xmax , respectively. The random
value chosen from a normal distribution is represented by rand when it
is in the range [0,1]. The EHO algorithm was evaluated for its perfor-
mance in a number of standard set functions as well as in medical
diagnosis, where it demonstrated improved accuracy. In this particular
investigation, the EHO algorithm is utilized for the purpose of DBN
parameter optimization. The outcome of the DBN model is based on the
weights and the biases of the layers of the network that came before it. In
contrast to other optimization techniques, EHO does not use the in-
dividuals who came before them in later stages of the updating process.
EHO is a swarm-inspired method that deals with global optimization
tasks that are considered by clan update and searching actions. EHO is
an acronym for Enhanced Hybrid Optimization. Because EHO does not
engage in relaxing, it is less susceptible to the effects of noise. They
execute better in situations that are limited and optimized for them.
Important aspects of EHO include a high convergence rate along with
low localization error rates, all while reducing the amount of time
needed for execution. The algorithm is capable of directly addressing
non-convex ML issues.

3.5. Fine tuning of DBN

In most RBMs, the four fundamental parameters that are adjusted are
the learning rate, the momentum weight, the hidden units, and the
weight decay. Due to the complexity and divergence of the problem, the
use of typical methods to compute the error function is considered an
NP-hard task. In order to resolve this issue, metaheuristics have been put
to use. In this work, the EHO method is applied in order to perfect the
DBN training by adjusting the parameters in minute detail. In this case,
the parameters that have been set are as follows: η = [0.1,0.9], n =
[5100], α = [0.00001,0.01]. and λ = [0.1,0.9], In order to accomplish
what needs to be done, there needs to be a fitness function that directs
the search process to the most appropriate responses. The fitness func-
tion that is used is referred to as mean squared error (MSE). The equa-
tion for this function, which determines the amount of deviation that
Fig. 5. Flowchart of EHO exists between the actual value and the target value, is provided in (5).

247
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

1∑ N ∑ D
( )2 each image patch, furthermore merged the individual tiles back together
MSE = Dj (i) − Yj (i) (5) to form the original image. This stage of stitching can lead to devising
T j=1 i=1
the borders of each individual patch, as demonstrated in the last two
In most RBMs, the four fundamental parameters that are adjusted are rows of Fig. 2. In order to solve this issue, the proposed system utilized a
the learning rate, the momentum weight, the hidden units, The EHO Conditional Random Field (CRF) technique in its entirety. Krahenbuhl
searches for a set of DBN parameters that results in the lowest possible and Koltun [25] were the first to suggest using this method to efficiently
MSE value. Dj (i) symbolizes the value that can be found in the jth unit of compute picture segmentations. It exhibited the ability to both capture
the DBN’s output layer at the period ‘t’, and Yj (i) indicates the jth tiny edge information as well as make use of dependencies that are
feature of the chosen value. The procedure is carried out repeatedly up long-range. After that, Chen et al. [26] implemented this strategy into
until the point where it can be stopped for good. The following are the CNNs as an additional processing step so that they could better analyze
1
steps involved in EHO’s optimization process and Fig. 5 show the images. P(I, X) = z(I) exp(− E (I, X)) is the formula that describes a CRF
flowchart of EHO: (I, X), and it is important to note that X is defined across the entire
picture as {x1 , x2 , … xN } is the total numeral of pixels, and xi is the label
1. Initialize the population and set the EHO parameters in the appro- that corresponds to the pixel that is numbered i. The energy function is
priate places. incorporated into the technique.
2. Determine the distinct fitness value of the DBN using the RMSE ∑ ∑ ( )
metric, taking into account both the learning rate and the total nu- E (I, X) = θi (xi ) + θi xi , xi,j (6)
meral of batch learning iterations. Determine the best possible per- i i,j

son to use.
where the first term is the unary latent and the 2nd term is the pairwise
3. Determine if the termination condition has been satisfied; if it has,
latent. The unary latent is demarcated as θi (xi ) = − log P(xi ), where
the iteration should be finished, and the result should be produced;
P(xi ) is the label obligation probability at pixel i as determined by the
otherwise, proceed to the following step.
segmentation head in the GNH. The unary potential is defined as the
4. Make necessary adjustments to each individual position. Reset the
difference between these two values. The pairwise potential is defined as
individuals’ values to their default values elsewhere the lower and ( )∑K ( )
upper bounds. the following equation: θi,j = μ xi , xj m=1 ω.k
m
fi , fj , where
( )
5. Begin a novel iteration by modifying the ideal individual using the μ xi , xj = 1 if xi ∕
= xj . Each km represents the Gaussian kernel, which is
previous results. weighted by a learnable parameter known as m and is dependent on
features that have been retrieved from pixels i and j and are designated
by the letter f. In the kernels, make use of bilateral position and color
Algorithm 1. optimal parameter selection of DBN by using EHO terms, following the example given in [46].

( ⃦ ⃦2 ( ⃦ ⃦)
⃦ ⃦ ⃦ ⃦ ) ⃦ ⃦
iii. Fully Connected Conditional Random Field Postprocessing ⃦pi− pj ⃦ ⃦Ii − Ij ⃦2 ⃦pi− pj ⃦
ω1 exp − − + ω2 exp − (7)
2σ 2α 2σ 2β 2σ2γ
In most RBMs, the four fundamental parameters that are adjusted are
the learning rate, the momentum weight, and the hidden units, once where p stands for the pixel’s location and I for the intensity of the
predictions were made using the suggested CDBN-EHO technique on pixel’s color. As a result, the first kernel term ensures that neighboring

248
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

pixels of a similar color are placed in the same class, whilst the second are used to evaluate the performance of segmentation findings. This
kernel term gets rid of small patches that are isolated from one another. allows us to make This suggested technique comparable with earlier
The “scale” of the Gaussian kernels, which were determined during the work. The following is an explanation of how these metrics are defined.
trial using an empirical method, is controlled by the hyperparameters Assume that furthermore have the results of the segmentation known as
σ α , σβ and σ γ , respectively. In the subsequent parts of this study, for the s f, the ground truth label known as tl, and a pixel-wise confusion matrix
sake of convenience, furthermore will refer to fully linked CRF as CRF. known as Cm. Here, Cmi,j represents the number of pixels that were
labeled as tli and were predicted to be sfj . The mIOU is defined as the
4. Experimental results and discussion mean of all of the individual Jaccard coefficients, which are denoted by
the notation Ji , for each class tli . Here makes use of the definition of the
In this part, furthermore will provide a quick overview of the pro- Jaccard index in order to derive Ji from the confusion matrix Cm.
posed CDBN-EHO experiment design, then present a number of experi-
TruePositive Cmi,i
mental data to validate the suggested design for the epithelial cell Ji = = (8)
TruePositive + False Positive + False Negative tpi + Pi − Cmi,i
identification and Gleason grading operations. The consequences of the
technique’s instance segmentation were translated into the semantic ∑
where tpi = Cmi,j denotes the total numeral of pixels with label tli . Pj =
segmentation outcomes by selecting the instance class with the highest j=1
∑
probability at each pixel location. This was done so that the new findings Cmi,j denotes the number of pixels predicted as f j. The mIOU is then
could be easily compared to the work that had been done previously. i

The suggested method is analyzed, and its performance is compared to specified by

that of the already established routes R–CNN [27], CNN [9], and DNN N
1∑
[8]. This section employs the conventional metrics: OPA (Overall Pixel J= Ji (9)
N
Accuracy), Standard Mean Accuracy (SMA), mean Intersection Over
Union (mIOU), in addition recall, accuracy, precision, and F-measure where N is the number of classes. The OPA is defined as

Fig. 6. Proposed CDBN-EHO technique results among existing technique.

249
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

∑
i Cmi,i
“stroma”, “benign”, and “high-grade” classifications, among these four
OPA = ∑ ∑ (10)
i Cmi,i categories. In spite of this, it only receives a “low-grade” IOU score of
j
71.54 %. This is because “low-grade” glands can have a wide range of
The SMA is well-defined as appearances from one another. The glands that are “lowgrade” might
vary in size and shape, although they are typically long and/or angular
1 ∑ Cmii in shape. The majority of them are microglandular, however some of
SMA = ∑ (11)
N i j Cmij them might be anywhere from medium to enormous in size. This di-
versity in size and shape is very clear in the second column of Fig. 6,
For the purpose of extracting features from the input pathological
where “low-grade” glands are represented by the green hue. This vari-
image, this section made use of a ResNet [28] in the suggested
ance can be plainly observed.
CDBN-EHO technique. Both the RPN and the GNH make use of a
Fig. 7 depicts the results obtained by the suggested technique after
structure known as a feature pyramid network (FPN) [29], which in-
the ENH and CRF were included. It demonstrates that the ENH provides
volves the substitution of feature pyramids for single-scale feature maps.
a significant improvement to the performance of the segmentation. This
The FPN creates feature pyramids with the format {P2 ….P6 }., as
is primarily due to the fact that This technique settings need us to make a
described in Ref. [29]. At each individual feature pyramid, here placed a
choice between the accuracy of This objectless prediction and the ac-
unique scale anchor that corresponds to a potential region of interest for
curacy of This segmentation. If Furthermore want This system to have a
the RPN. After that, the RPN is trained with the parameters that are
high accuracy that minimizes the failure to detect possible epithelial
shared across all levels of the feature pyramid. Furthermore allocate
regions, then Furthermore need to lower the detection threshold when
each Region of Interest (ROI) on the input picture to the GNH network,
ENH is not present because that will allow us to have high precision. This
which has dimensions of width wi and height ht, to the feature pyramid
will provide us with a technique that is meant to forecast epithelial cells
Pyk using the following formula:
⌊ (√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅)⌋
Pyk = k0 + log2 wi*ht/224 (12) Table 2
Overall performance evaluation among all cancer segmentation schemes.
If the scale of the RoI is reduced (for example, to 1/2 of 224), the Performance metrices Proposed CDBN- Path CNN DNN
equation eqn. (12) suggests that the data should be mapped onto a level (%) EHO R–CNN
with a higher resolution (for example, k = 3). The information that the
Accuracy 95.86 94.75 92.64 93.41
technique needs to send into the GNH is extracted from each RoI by the Precision 97.32 96.67 94.22 92.60
technique using this technique, which is done on a similar scale. The Recall 97.88 96.77 95.19 94.57
outcomes of the projected and existing technique are illustrated in Fig. 6. F-measure 97.95 96.77 95.19 94.08
The suggested technique yields mIOU scores of 81.32 %, SMA scores of mIOU 81.32 77.56 74.22 72.60
SMA 89.78 87.58 83.53 82.43
89.78 %, and OPA scores of 90.40 % across all four classes. Path R–CNN OPA 90.40 88.9 87.6 84.08
has a performance that is considered to be above average in the

Fig. 7. The effectiveness of incorporating the ENH and CRF into the proposed CDBN-EHO structure.

250
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

Fig. 8. PC segmentation accuracy prediction among methods.

Fig. 10. PC segmentation recall prediction among methods.

more frequently even in an image that is full of stroma; as a result, the

careful optimization of its parameters.
performance will be drastically lowered. The first two rows of Fig. 7 are
Fig. 8 depicts the accuracy of PC segmentation achieved by the
a good example of this phenomenon. Furthermore, can observe that the
proposed CDBN-EHO algorithm and evaluates its performance in com-
technique has a tendency to anticipate ROIs in vast sections of stroma by
parison to that of other current segmentation methods such as route
looking at the very last column. As a result, the proposed technique
R–CNN, CNN, and DNN. Due to a reduction in the amount of computing
arrives at the conclusion that the ENH is essential to the success of
time required, the accuracy of the projected technique improved along
achieving good performance in the proposed system. In the Supple-
with the number of photos that were used. Based on the findings, it can
mentary Information, further justifications and benefits of the ENH are
be deduced that the proposed CDBN-EHO algorithm yields high accu-
explored. As can be seen in the last two rows of Fig. 5, the results that
racy results of 95.86 %, whereas the accuracy yielded by other classi-
have been proposed utilizing the CRF reveal that the addition of the
fiers, such as R–CNN, CNN, and DNN, respectively, is 94.75 %, 92.64 %,
approach helps remove the unnatural borders that are caused by
and 93.41 %. The performance of DBN has improved as a result of the
stitching. The unnatural boundaries that were produced as a result of the
optimal prediction of parameter value.
stitching procedure are denoted by the red arrows in the picture (Rows 3
Fig. 9 displays the precision of PC segmentation achieved by the
and 4). After the CRF has been post-processed, furthermore see that
proposed CDBN-EHO algorithm, which was tested in comparison to
these artificial borders have been erased. Additionally, the CRF con-
other current segmentation schemes such as route R–CNN, CNN, and
tributes to a moderate improvement in mIOU.
DNN. Due to a reduction in the amount of computing time required, the
The total performance of the proposed CDBN-EHO in comparison to
precision of the proposed strategy improved along with the number of
other current systems such as R–CNN, CNN, and DNN is displayed in
photos that were used. Based on the findings, it can be deduced that the
Table 2. It is concluded that the suggested CDBN-EHO achieved great
suggested CDBN-EHO algorithm achieves high precision results of
performance results in comparison to existing schemes due to their ca-
97.32 %, but other classifiers, such as the R–CNN, CNN, and DNN
pacity to process huge datasets and need less computational time. The
techniques, only achieve precision values of 96.67 %, 94.22 %, and
performance of DBN has been boosted across the board thanks to the
92.60 % values respectively. The study comes to the conclusion that the

Fig. 9. PC segmentation precision prediction among methods. Fig. 11. PC segmentation F-measure prediction among methods.

251
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

suggested CDBN-EHO algorithm achieves high F-measure results of

97.95 %, whereas other classifiers, such as R–CNN, CNN, and DNN
techniques, generate F-measure values of 96.77 %, 95.19 %, and 94.08
% values respectively. It comes to the conclusion that research work is
more successful when paired up with other methods.
Fig. 12 depicts the PC segmentation mIOU of the proposed CDBN-
EHO, which was then evaluated in comparison to other current seg-
mentation schemes such as route R–CNN, CNN, and DNN. Because there
is less time spent computing, the mIOU of the proposed approach in-
creases in proportion to the number of photos that are being used. Based
on the findings, it can be deduced that the proposed CDBN-EHO algo-
rithm achieves high mIOU results of 81.32 %, whereas other classifiers,
such as R–CNN, CNN, and DNN techniques, only achieve mIOU values of
77.56 %, 74.22 %, and 72.60 %, respectively. The study comes to the
conclusion that the CDBN-EHO research method performs better when
compared to other methodologies.
Fig. 13 depicts the PC segmentation SMA of the proposed CDBN-
EHO, which was then evaluated in comparison to other current seg-
mentation methods such as route R–CNN, CNN, and DNN. Due to the
reduced amount of time spent computing, the SMA of the suggested
approach increases together with the number of photos that are being
Fig. 12. PC segmentation mIOU prediction among methods. used. Based on the findings, it can be deduced that the proposed CDBN-
EHO algorithm achieves high SMA results of 89.78 %, while other
CDBN-EHO research method performs better when compared to other classifiers, such as R–CNN, CNN, and DNN techniques, only achieve SMA
methodologies. values of 87.58 %, 83.53 %, and 82.43 % respectively. The study comes
The PC segmentation recall of the proposed CDBN-EHO is displayed to the conclusion that the CDBN-EHO research method performs better
in Fig. 10, and it is evaluated in comparison to other existing segmen- when compared to other methodologies. The proposed method achieves
tation methods such as route R–CNN, CNN, and DNN. Due to a reduction high SMA through a combination of innovative methodologies, effective
in the amount of computing time required, the recall of the suggested Bayesian inference, multi-task learning, optimized parameterization,
technique improved along with the number of photos that were used. and comprehensive evaluation. By introducing novel optimization
Based on the findings, it can be deduced that the suggested CDBN-EHO techniques like EHO-based Hyper-parameter CDBN and grading
method achieves high recall results of 97.88 %, but other classifiers, network head-optimized DBN, the model effectively captures relevant
such as the R–CNN, CNN, and DNN techniques, only achieve recall features and contextual information crucial for accurate histological
values of 96.77 %, 95.19 %, and 94.57 % values respectively. The study image analysis. Leveraging an extremely effective Bayesian inference
comes to the conclusion that the CDBN-EHO research method performs approach for fully linked CRF techniques ensures precise segmentation,
better when compared to other methodologies. contributing to the overall accuracy of the model. The multi-task
The PC segmentation F-measure of the proposed CDBN-EHO is dis- approach enables simultaneous consideration of multiple tasks,
played in Fig. 11, and it is evaluated in comparison to other current enhancing performance by leveraging shared contextual information
segmentation schemes such as path R–CNN, CNN, and DNN. Due to the across tasks. Additionally, optimized parameterization minimizes
decreased amount of time spent computing, the F-measure of the sug- overfitting and algorithmic sensitivity, ensuring robust performance
gested scheme increased along with the number of images as they were across diverse datasets and clinical scenarios. This comprehensive
added. Based on the findings, one may draw the conclusion that the evaluation using SMA reflects the model’s reliability and effectiveness in

Fig. 13. PC segmentation SMA prediction among methods. Fig. 14. PC segmentation OPA prediction among methods.

252
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

real-world clinical settings, particularly in tasks like prostate cancer CRediT authorship contribution statement
diagnosis and Gleason grading.
Fig. 14 depicts the PC segmentation OPA of the proposed CDBN- S. Angel Latha Mary: Investigation, Methodology, Project admin-
EHO, which was then evaluated in comparison to other existing seg- istration. S. Siva Subramanian: Conceptualization, Data curation,
mentation methods such as route R–CNN, CNN, and DNN. The OPA of Formal analysis. G. Priyanka: Validation, Visualization. T. Vijayaku-
the proposed technique grew when there was less time spent computing, mar: Writing – original draft, Writing – review & editing. Suganthi
which led to an increase when the number of photos was increased. Alagumalai: Resources, Software, Supervision.
Based on the findings, it can be deduced that the suggested CDBN-EHO
algorithm achieves high OPA results of 90.40 %, whereas other classi- Declaration of competing interest
fiers, such as R–CNN, CNN, and DNN algorithms, generate OPA values of
88.9 %, 87.6 %, and 84.08 % values respectively. It comes to the The authors declare the following financial interests/personal re-
conclusion that research work is more successful when paired up with lationships which may be considered as potential competing interests:
other methods. The high OPA achieved by the proposed method can be Suganthi Alagumalai is currently employed by JPMorgan Chase, where
attributed to several key factors. The model effectively delineates she serves as an Agility Lead role within the Cyber Security division.
boundaries and structures within histological images through precise
segmentation techniques. By accurately identifying pixel-level details, References
the model ensures that each pixel is correctly classified, leading to high
pixel accuracy. The proposed method utilizes advanced DL architec- [1] PC: The Leading Cancer in India.
tures, such as CDBNs, which are optimized for pixel-level classification [2] S. Ghai, M.A. Haider, Multiparametric-MRI in diagnosis of PC, Indian J. Urol: IJU: j.
Urol. Soc. India 31 (3) (2015) 194.
tasks. The architecture is designed to capture intricate patterns and [3] R. Loffroy, O. Chevallier, M. Moulin, S. Favelier, P.Y. Genson, P. Pottecher,
features within the images, contributing to accurate pixel-wise pre- L. Cormier, Current role of multiparametric magnetic resonance imaging for PC,
dictions. The evaluation of OPA provides a comprehensive assessment of Quant. Imag. Med. Surg. 5 (5) (2015) 754.
[4] Rights, P. B. O., Chipko, C. R., Judy, G. D., Liebman, L. R., Pennington, J. D.,
the model’s performance at the pixel level, accounting for both true
Siddiqui, O., … & Links, R. O. A. Radiation Therapy for PC.
positive and true negative predictions. This holistic evaluation metric [5] D.A. Barron, D.R. Rowley, The reactive stroma microenvironment and PC
ensures that the model’s segmentation results are accurate and reliable, progression, Endocr. Relat. Cancer 19 (6) (2012) R187–R204.
[6] S. Iqbal, G.F. Siddiqui, A. Rehman, L. Hussain, T. Saba, U. Tariq, A.A. Abbasi, PC
further validating its effectiveness in histological image analysis tasks.
detection using DL and traditional techniques, IEEE Access 9 (2021) 27085–27100.
[7] 1. Saqib Iqbal, Ghazanfar Farooq Siddiqui, Amjad Rehman, Lal Hussain,
5. Conclusion Tanzila Saba, Usman Tariq, Adeel Ahmed Abbasi, PC detection using DL and
traditional techniques IEEE Access 9 (2021) 27085–27100.
[8] Yuchun Li, Mengxing Huang, Yu Zhang, Jing Chen, Haixia Xu, Gang Wang,
In conclusion, our paper introduces the innovative CDBN-EHO Wenlong Feng, Automated gleason grading and gleason pattern region
framework alongside a grading network head-optimized deep belief segmentation based on DL for pathological images of PC, IEEE Access 8 (2020)
network technique for multi-task prediction in histological image anal- 117714–117725.
[9] Davood Karimi, Guy Nir, Ladan Fazli, Peter C. Black, Larry Goldenberg, Septimiu
ysis. We leverage advanced methodologies such as Bayesian inference E. Salcudean, DL-based gleason grading of PC from histopathology images—role of
for segmentation using fully linked Conditional Random Field tech- multiscale decision aggregation and data augmentation, IEEE J. Biomed Health Inf.
niques, with pairwise boundary capacities determined by a linear 24 (5) (2019) 1413–1426.
[10] Lourdes Duran-Lopez, Juan P. Dominguez-Morales, Antonio Felix Conde-Martin,
mixture of Gaussian kernels. Our approach demonstrates the potential to Saturnino Vicente-Diaz, Alejandro Linares-Barranco, PROMETEO: a CNN-based
enhance performance by incorporating more contextual information computer-aided diagnosis system for WSI PC detection, IEEE Access 8 (2020)
through multi-task learning. Overall, this research contributes to 128613–128628.
[11] Ohad Kott, Drew Linsley, Ali Amin, Andreas Karagounis, Carleen Jeffers,
advancing histological image analysis and lays the groundwork for the
Dragan Golijanin, Thomas Serre, Boris Gershman, Development of a DL algorithm
development of more accurate, efficient, and interpretable diagnostic for the histopathologic diagnosis and Gleason grading of PC biopsies: a pilot study,
tools for improving patient care in domains such as cancer diagnosis and Eur. Urol. focus 7 (2) (2021) 347–351.
[12] Dallin Busby, Ralph Grauer, Krunal Pandav, Akshita Khosla, Parag Jain,
treatment. This was the case in terms of accuracy, precision, recall, f-
Mani Menon, G. Kenneth Haines III, Carlos Cordon-Cardo, Michael A. Gorin,
measure, 88.78 % SMA, 79.56 % mIOU, and 89.40 % OPA, to name a Ashutosh K. Tewari, Applications of artificial intelligence in PC histopathology, in:
few metrics. In the future, the identification of PC systems will be Urologic Oncology: Seminars and Original Investigations, Elsevier, 2023.
improved by employing the feature selection and ranking phase with [13] Misbah Ahmad, Imran Ahmed, Ahmed Ouameur Messaoud, Gwanggil Jeon,
Classification and detection of cancer in histopathologic scans of lymph node
various hybrid optimization methods. This will be done in order to sections using convolutional neural network, Neural Process. Lett. (2022) 1–16.
improve accuracy. In addition, granular computing will be included into [14] Zaneta Swiderska-Chadaj, Thomas de Bel, Lionel Blanchet, Alexi Baidoshvili,
DNNs in the near future in order to greatly rise the pace of calculation. Dirk Vossen, Jeroen van der Laak, Geert Litjens, Impact of rescanning and
normalization on convolutional neural network performance in multi-center,
Moving forward, several avenues of future work are identified. Firstly, whole-slide classification of PC, Sci. Rep. 10 (1) (2020) 14398.
we aim to further optimize our framework to mitigate algorithmic sen- [15] Abhinav Kumar, Sanjay Kumar Singh, K. Lakshmanan, Sonal Saxena,
sitivities and enhance generalization across diverse datasets and clinical Sameer Shrivastava, A novel cloud-assisted secure deep feature classification
framework for cancer histopathology images, ACM Trans. Internet Technol. 21 (2)
scenarios. Additionally, efforts will be directed towards improving the (2021) 1–22.
interpretability of our models to foster greater trust among clinicians. [16] Andrea Chatrian, Korsuk Sirinukunwattana, Clare Verrill, Jens Rittscher, Towards
The computational complexity and trade-offs of different methods for the identification of histology-based subtypes in PC, in: 2019 IEEE 16th
International Symposium on Biomedical Imaging (ISBI 2019), IEEE, 2019,
histopathological image analysis in prostate cancer diagnosis are crucial
pp. 948–952.
considerations. Techniques like the Elephant Herding Optimization- [17] N. Ing, et al., Semantic segmentation for PC grading by convolutional neural
based Hyper-parameter Convolutional Deep Belief Network (CDBN- networks, Proc. SPIE 10581 (Mar. 2018) 105811B.
[18] A. Gertych, et al., Machine learning approaches to analyze histological images of
EHO) may offer breakthrough results but could potentially require more
tissues from radical prostatectomies, Comput. Med. Imag. Graph. 46 (Dec. 2015)
memory space and time compared to other methods like R–CNN, CNN, 197–208.
and DNN. Evaluating these trade-offs is essential for selecting the most [19] Stephen O’Hara, Bruce A. Draper, Introduction to the Bag of Features Paradigm for
suitable approach for accurate and timely diagnosis, considering the Image Classification and Retrieval, 2011 arXiv preprint arXiv:1101.3354.
[20] Tapas Kanungo, David M. Mount, Nathan S. Netanyahu, Christine D. Piatko,
urgency of detecting and treating prostate cancer, the second highest Ruth Silverman, Angela Y. Wu, An efficient k-means clustering algorithm: analysis
cause of cancer-related deaths in men globally. and implementation, IEEE Trans. Pattern Anal. Mach. Intell. 24 (7) (2002)
881–892.
[21] G. Csurka, C.R. Dance, L. Fan, J. Willamowski, C. Bray, Visual categorization with
bags of keypoints, in: Workshop on Statistical Learning in Computer Vision, 2004,
pp. 1–22.

253
S. Angel Latha Mary et al. International Journal of Intelligent Networks 5 (2024) 241–254

[22] T. Tatiana, O. Francesco, C. Barbara, CLEF2007 image annotation task: an SVM- [26] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A.L. Yuille, DeepLab: semantic
based cue integration approach, in: Proceedings of Image CLEF, 2007. image segmentation with deep convolutional nets, atrous convolution, and fully
[23] H. Lee, Y. Largman, P. Pham, H. Lee, P. Pham, Y. Largman, Unsupervised feature connected CRFs, IEEE Trans. Pattern Anal. Mach. Intell. 40 (4) (Apr. 2017)
learning for audio classification using convolutional deep belief networks, in: 22nd 834–848.
International Conference on Neural Information Processing Systems, ACM, [27] W. Li, J. Li, K.V. Sarma, K.C. Ho, S. Shen, B.S. Knudsen, C.W. Arnold, Path R-CNN
Vancouver, British Columbia, Canada, 2009, pp. 1096–1104. for PC diagnosis and gleason grading of histological images, IEEE Trans. Med.
[24] M. Nayak, S. Das, U. Bhanja, M.R. Senapati, Elephant herding optimization Imag. 38 (4) (2018) 945–954.
technique based neural network for cancer prediction, Inform. Med. Unlocked 21 [28] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
(2020) 100445. Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Jun. 2016, pp. 770–778.
[25] P. Krähenbühl, V. Koltun, Efficient inference in fully connected CRFS with [29] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, S.J. Belongie, Feature pyramid
Gaussian edge potentials, in: Proc. Adv. Neural Inf. Process.Syst., 2011, networks for object detection, Proc. CVPR 1 (2) (Jul. 2017) 4.
pp. 109–117.

254