Information Sciences 640 (2023) 119005

Contents lists available at ScienceDirect

Information Sciences
journal homepage: www.elsevier.com/locate/ins

Application of Kronecker convolutions in deep learning technique

for automated detection of kidney stones with coronal CT images
Kiran Kumar Patro a,∗ , Jaya Prakash Allam b,∗ , Bala Chakravarthy Neelapu c ,
Ryszard Tadeusiewicz d , U Rajendra Acharya e , Mohamed Hammad f ,g ,
Ozal Yildirim h , Paweł Pławiak i,j,∗∗
a
Department of Electronics and Communication Engineering, Aditya Institute of Technology and Management, Tekkali, AP-532201, India
b
Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Rourkela, Odisha-769008, India
c
Department of Biotechnology and Medical Engineering, National Institute of Technology Rourkela, Rourkela, Odisha-769008, India
d
AGH University of Science and Technology, Department of Biocybernetics and Biomedical Engineering, Krakow, Poland
e
School of Mathematics, Physics and Computing, University of Southern Queensland, Springfield, Australia
f
EIAS Data Science Lab, College of Computer and Information Sciences, Prince Sultan University, Riyadh 11586, Saudi Arabia
g
Department of Information Technology, Faculty of Computers and Information, Menoufia University, Shibin El Kom 32511, Egypt
h
Department of Software Engineering, Firat University, Elazig, Turkey
i
Department of Computer Science, Faculty of Computer Science and Telecommunications, Cracow University of Technology, Warszawska 24,
31-155 Krakow, Poland
j
Institute of Theoretical and Applied Informatics, Polish Academy of Sciences, Bałtycka 5, 44-100 Gliwice, Poland

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

Keywords: Kidney stone disease is a serious public health concern that is getting worse with changes in
Coronal CT scan diet, obesity, medical conditions, certain supplements etc. A kidney stone also called a renal
Image embossing calculus, is a hard buildup of urine minerals that form in the kidneys. Computed tomography
Kidney stone
(CT) is one of the imaging models used to identify kidney stones by clinical experts. Due to
Kronecker convolution
Pre-processing
the low resolution of these images, sometimes detecting kidney stones is tedious with the naked
eye, which may lead to false alarms. In this work, a computer-based diagnosis system with a
deep learning technique has been developed as a practical solution to aid clinicians in their
diagnosis. The traditional convolutional neural network (CNN)-based deep learning technology
can detect stones in the kidney. Still, it suffers from the performance and standard implementation
of the convolution operations in convolution layers. A Kronecker product-based convolution
technique is incorporated in the proposed deep learning architecture to reduce the redundancy
in feature maps without convolution overlapping. Our proposed method helps to make the
network more effective by extracting abstract and in-depth features from the input images. The
publicly available GitHub kidney stone CT scans are utilized to develop the proposed architecture.
Our automated model detected kidney stones with an accuracy of 98.56% utilizing CT images.
Our system is more effective than the most recent and cutting-edge techniques developed for
identifying kidney stones of any size, including the smallest ones.

Corresponding authors at: 𝑎 Department of Electronics and Communication Engineering, Aditya Institute of Technology and Management, Tekkali, AP-532201,
*
India and 𝑏 Department of Electronics and Communication Engineering, National Institute of Technology Rourkela, Rourkela, Odisha-769008, India.
** Corresponding author at: Department of Computer Science, Faculty of Computer Science and Telecommunications, Cracow University of Technology, Warszawska
24, 31-155 Krakow, Poland.
E-mail addresses: [email protected] (K.K. Patro), [email protected] (J.P. Allam), [email protected] (P. Pławiak).

https://doi.org/10.1016/j.ins.2023.119005
Received 9 January 2023; Received in revised form 19 April 2023; Accepted 21 April 2023
Available online 25 April 2023
0020-0255/© 2023 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

1. Introduction

Kidney stones, or renal calculi, are one of the most common urological illnesses, affecting an estimated 1-13% of the world’s
population [1]. Kidney stone formation, or renal calculus, is defined by the formation of crystals in the urine due to drug concen-
tration or hereditary vulnerability. Stones are pebble-shaped structures that develop in the kidney due to the solid concentration
or crystal aggregation of mineral components in the diet. Even infants are prone to developing kidney stones; nevertheless, most
cases go undiagnosed unless the patient displays acute abdominal pain or an odd colour in their urine. However, the most crucial
elements that lead to the formation and development of a kidney stone are a person’s diet, lifestyle, and the effects of acquired
metabolic abnormalities. Kidney stones can be categorized based on their chemical makeup, and their location [2]. They are calcium
oxalate, struvite, uric acid, and cystine stones according to their respective chemical makeups. Nephrolithiasis refers to kidney stones,
ureterolithiasis indicates stones that form in the ureter, and cystolithiasis is a bladder stone.
Moreover, patients with kidney stones often experience symptoms such as fever, discomfort, and nausea, which can be mistaken
for the symptoms of other diseases. Kidney stones can also limit the kidneys’ ability to dilate and perform their normal functions.
Thus, it is crucial to diagnose kidney stones, especially in their early stages, to provide the appropriate medical treatment to pa-
tients. Biomedical imaging is now crucial in both clinical and scientific research. Medical imaging creates visual representations
of interior organs for a clinical study. The most reliable methods of diagnosing kidney stones are imaging techniques that rely on
computer technology, such as ultrasonography, computed tomography (CT), and X-rays obtained via intravenous pyelogram (IVP).
Most hospitals prefer to screen for kidney stones using CT scans due to the convenience of the three-dimensional images. CT tech-
nology advancements are essential for clinicians and patients due to their ease of use and effectiveness in detecting kidney stones in
both asymptomatic and symptomatic patients [3]. Furthermore, CT imaging can determine the size and location of the stone, which
enables the classification of the risk of spontaneous stone passage without requiring surgical intervention. Some samples of coronal
CT scans with kidney stones and normal images are shown in Fig. 1.
Currently, various medical imaging domains have leveraged the combination of computer vision and deep learning techniques.
Deep learning models have demonstrated successful applications in several fields, including medical imaging segmentation [4],
classification [5], and lesion detection [6]. Various imaging modalities like magnetic resonance imaging (MRI), CT, and X-ray images
have been used to train deep learning (DL) models that are accurate and robust enough to aid doctors in diagnosing conditions like
cardiac abnormalities, COVID-19, prostate cancer, brain tumors, breast cancers, and skin cancers. Using deep learning and computer
vision techniques, Poyraz et al. [7] proposed a model for the automated diagnosis of brain diseases based on pre-existing exemplars.
The features used in the classification task were derived using a pre-trained MobilNetV2, and a support vector machine (SVM) was
used for classification. To detect and categorize uncommon skin diseases from medical images, Ravi et al. [8] suggested a deep
learning-based feature fusion ensemble approach. The skin disease prediction takes place in the first stage of the ensemble meta-
classifier, and the classification is performed in the second stage using the results of the first stage as features. Recently, authors
in [9] developed a deep learning model to identify and detect childhood pneumonia by employing chest X-rays as the primary
diagnostic tool (CXR). To achieve successful outcomes, a transfer learning-based deep feature fusion strategy for pediatric pneumonia
classification was applied. Moreover, several multichannel deep-learning techniques for disease detection and classification have been
developed [10]. Each model’s attributes are combined and fed into a stacked ensemble learning classifier to ensure accurate disease
detection.
Medical image processing has been extensively used by numerous researchers to study renal issues and kidney stone diseases
[11]. Shi et al. [12] proposed a fully automatic segmentation method for clinical ultrasonography kidney images. To extract high-
level image features, pre-trained deep neural networks were used by the authors for the categorization of natural images. In [13],
the authors introduced a deep learning-based metric, called the pathological estimated fibrosis score (PEFS), to identify patients
with chronic renal failure. A deep learning technique was employed by combining patient-specific histological objects with clinical
phenotypes such as serum creatinine, nephrotic proteinuria at biopsy, chronic kidney disease (CKD), and renal survival for 1 to
5 years. Abdullah et al. [14] proposed the early detection of CKD using machine learning techniques. Missing data values were
recalculated based on the average of relevant attributes to enable experiments to be carried out. Following this, artificial neural
networks (ANN) and SVM were employed for the classification task.
Yildirim et al. [15] have recently developed a computer-assisted diagnostic technique for detecting kidney stones in coronal CT
scans using deep learning methods. The authors utilized recent advancements in artificial intelligence to analyze CT image features,
allowing them to detect even the smallest kidney stones with high accuracy. In another study, Baygin et al. [16] introduced ExDark19,
a transfer learning-based image classification method for automated kidney stone detection using CT images. They used the iterative
neighbourhood component analysis (INCA) to choose feature vectors containing the most relevant information, which were then
fed into a k-nearest neighbour (KNN) classifier for stone identification. The authors achieved an accuracy of 99% with a ten-fold
cross-validation strategy. Table 1 summarizes various studies that have employed computer-assisted tools for kidney stone detection.
CT scans are a complex and time-consuming method for detecting small kidney stones. Therefore, a computerized system can
support medical professionals in the precise diagnosis of kidney stones. This study presents a strong model that combines deep neural
networks with coronal CT image features for the accurate detection of kidney stones. The developed model can assist radiologists in
making precise diagnoses of kidney stones. The significant contributions of the study are as follows:

• The authors have developed a dataset containing cropped images of coronal CT scans. The significant component of the kidney
image (an area of interest region) is extracted from the entire CT scan, which enables a more accurate kidney stone diagnosis.

2
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 1. Coronal Kidney CT Scans: (a), (b), (c), (d), (i), (j), (k), and (l) are the samples of coronal CT Scans with kidney stone images and (e), (f), (g), (h) are the normal
healthy images.

Table 1
Summary of recent studies conducted on automated detection of kidney stones.

Author Renal disease Modality Technique Performance

Black et al., 2020 [17] Kidney stones composition Digital stone images Deep convolutional neural 83.2% overall accuracy for
(63 images) network with computer vision detecting four types of stone
algorithm composition
Martin et al., 2018 [18] Urethral stones High-resolution CT scans Convolutional Neural 0.996 AUC and 3.69 False
(465 images) Network (CNN) Positive rate per scan
Aziyus et al., 2020 [19] Urinary Stones Micro-CT scans Convolutional Neural 99.59% Test accuracy
(2430 Images) Network (CNN)
Xiang et al., 2019 [20] Urolithiasis Urine sediment microscopy ResNet50 74% Recognition rate
images (80 subjects)
Kazemi et al., 2018 [21] Kidney stones Numerical features from Ensemble Learning model 97.1% accuracy
936 patients
Chaitanya et al., 2019 [22] Chronic Kidney Disease Kidney MRI scan (27 scans Machine Learning classifiers 78% accuracy
(CKD) CKD and 6 Normal) with Gravitational Search (ANN with GSA)
Algorithm
Suresh et al., 2021 [23] Kidney stones Ultrasound images KUB Image processing Techniques 92.57% stone detection rate
(Pre-processing, segmentation
and Morphological approach)
Skounakis et al., 2013 [11] Tumors, Stones and cysts MRI scans (552 images) Template-based Techniques 96.1% average accuracy
Kahani et al., 2020 [24] Urinary stones X-ray KUB images LASSO with ML classifiers 96% accuracy
Liu et al., 2022 [25] Urolithiasis X-ray KUB images Residual Neural Network 97.70% accuracy
(355 images)
Kobayashi et al., 2021 [26] Urinary tract stones X-ray KUB images CNN based models 0.752 F-score
(1017 patients)

• The authors proposed a new Kronecker convolution method to remove the redundancy in the feature vectors and improve the
detection accuracy.
• Designed a simple model by incorporating different convolution methods to increase the model’s efficiency without increasing
the network’s depth. Therefore, authors are able to design a less complex and more efficient system for identifying kidney stones.
• Kronecker convolution enables us to concentrate on each pixel of data in the CT scan. It operates on each pixel of data in the CT
scan image to ensure an accurate diagnosis of even the smallest kidney stones.
• An ablation investigation was also performed to verify the significance of the proposed deep Kronecker neural network (DKN)
architecture.

The remaining sections of the paper are organized as follows. The data used for the proposed method is provided in Section 2.
The suggested DKN model’s architecture, along with pseudocode, is presented in Section 3, along with several intermediate image

3
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 2. Proposed methodology for automated kidney stone detection.

pre-processing steps. The results of the proposed method are presented in Section 4, and these results are compared with the state-
of-the-art techniques in this section. In Section 5, a DKN ablation study is conducted. Finally, the paper concludes in Section 6.

2. Materials

The dataset utilized in this work is obtained from the kidney stone detection repository on GitHub [15]. The repository contains
a diverse set of coronal CT scans acquired from multiple institutions and scanners to develop robust and accurate algorithms for
automated kidney stone detection. The scans in the dataset were acquired in Digital Imaging and Communications in Medicine
(DICOM) format and pre-processed to remove any personal identification information about the patient. The dataset consists of CT
scans from male and female patients, ranging in age from 18 to 80 years old, kidney stone patients. Radiologists or urologists who
reviewed the scans confirmed the diagnosis of kidney stones. At least two radiologists reviewed each CT scan in the dataset to ensure
the accurate detection and annotation of kidney stones. This database is made publicly available in the following link: (https://
github.com/yildirimozal/Kidney_stone_detection). Various imaging modalities can detect kidney stones, including ultrasonography,
KUB (kidney, ureters and bladder), CT, MRI, etc. In comparison to other diagnostic methods, CT has the highest sensitivity for
detecting kidney stones, and almost all stones can be seen effectively in a CT scan [27]. Researchers [15] acquired 500 coronal CT
scan images from individuals who had been admitted to the Elazig Fethi Sekin City Hospital in Turkey for the treatment of urinary
system stone disorders. In this study, 268 participants reported positive results on a stone test and 165 reported were normal (without
kidney stones). There are 1799 coronal CT scans in the database (1009:normal subjects and 790:patients with kidney stones). All
scans were taken with the patient lying face up on a single scanner without the use of contrast [15].

3. Proposed methodology for kidney stone image classification

The objective of this work is to create a model that can minimize the number of cases overlooked by doctors during the CT
image-based screening for kidney stones. A DKN-based model is developed using a combination of radiographic image processing
and deep learning techniques for the classification of kidney stones from low-contrast coronal CT scans. The methodology for this
study involved three main stages: dataset collection, database pre-processing, and classification using a customized DKN, as depicted
in Fig. 2.

3.1. Pre-processing of the input raw CT scan images

The database comprises kidney stones and normal coronal CT pictures in .png format of various sizes. To extract the region of
interest from an image, further distinct image-enhancing techniques such as image embossing, region identification, and cropping
were developed. Image embossing is a purely digital method of visual enhancement in which the visibility of each image pixel is
controlled by its proximity to the image’s edges. The embossing process involves transforming a two-dimensional image into a three-
dimensional one. Applying the emboss filter typically results in creating an image comparable to the original image. Still, it appears
to have been embossed onto a sheet of paper or piece of metal [28]. This method helps to obtain the highest possible contrast to be
achieved without affecting the spatial resolution.
Enhancing the edges is, consequently, the primary benefit of utilising this method. After applying an embossing filter to the input
coronal CT scan, relevant regions are highlighted in contrast to the background. Fig. 3 depicts the resulting embossed image, which

4
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 3. (a) Original image, and (b) embossing the image.

Fig. 4. Region of interest kidney stone detection labelled image.

Fig. 5. Cropped kidney stone detection labelled image based on ROI.

has a resolution of 1000×600. After embossing, significant areas are presented to show the region of interest to the model so that it
may focus more on detecting kidney stones.
In accordance with the opinion of a specialized medical practitioner, the portion of the kidney stone that has been diagnosed
is displayed with a red arrow and a green box shown in Fig. 4. In addition to this, the image was cropped based on the region of
interest in order to facilitate more accurate kidney stone detection. The cropping process produced an image with dimensions of
550x335 pixels, shown in Fig. 5. To attain good performance, deep neural networks require a significant amount of training data.
Image augmentation is synthetically creating different training images by combining various processing techniques [11,12,17]. In
this work, several image augmentation procedures, such as rotating the image by 30 degrees, flipping it horizontally, shearing it by
0.3, and zooming it by 0.3, were carried out to improve the performance of deep neural networks.

3.2. Proposed deep learning model for kidney stone detection

Deep learning is a sub-field of artificial intelligence that has successfully revealed impressive progress in various fields. Deep CNN
plays an important role in medical imaging systems. It enables radiologists to make more accurate diagnoses by offering a quantitative
analysis of suspicious areas, and an improved workflow [29]. In this work, authors developed deep Kronecker convolutional neural
network [30] architecture for kidney stone detection utilizing image features from intermediate stages. Initially, the Kronecker
convolution tree structure is utilised in [31] for semantic segmentation. In this work, for the first time, we developed an end-to-end
classification system based on Kronecker convolution. Segmentation and classification are two common tasks in image processing
and computer vision. In general, segmentation refers to the process of dividing an image into different regions or segments, while
classification refers to the task of predicting the class or label of an input. Segmentation is typically used to identify different objects

5
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 6. Graphical illustration of Kronecker and standard convolution operation on 2×2 image.

or regions within an image and can be used for tasks such as object detection and image annotation. This is often done by dividing
the image into a grid of pixels and assigning each pixel to a particular class or label. Classification, on the other hand, is the task of
predicting the class or label of an input based on some learned rules or patterns. Kronecker convolution layers are utilized in this
work to reduce the overall parameters of the network and complex computations. The authors introduced Kronecker convolution
instead of regular convolution to exploit the local structures within the pixel. The Kronecker convolution of two polynomials f(x)
and g(x) is denoted 𝑓 (𝑥) ⊗ 𝑘𝑔(𝑥) and is defined in eq. (1) as follows:
𝑛
∑
𝑓 (𝑥) ⊗ 𝑘𝑔(𝑥) = 𝑓𝑖 ∗ 𝐺 𝑖 (1)
𝑖=0

where 𝑓𝑖 is the coefficient of the 𝑥𝑖 term in f(x), G is a matrix representation of the finite field element x, and n is the degree of the
polynomial.
In a standard CNN, the output of a convolutional layer with input tensor X and weight tensor W is shown in eq. (2) is given by:

𝑜𝑢𝑡𝑝𝑢𝑡 = X ∗ W + b (2)
where “*” is the standard convolution operator and b is a bias term.
In a Kronecker convolutional neural network, the output in eq. (3) is instead computed using the Kronecker product (also called
the tensor product) of the input and weight tensors:

𝑜𝑢𝑡𝑝𝑢𝑡 = 𝑋 ⊗ 𝑊 + 𝑏 (3)
where ⊗ is the Kronecker product operator.
The Kronecker product of two matrices, A and B, shown in eq. (4), is defined as:

𝐴 ⊗ 𝐵 = [𝑎(𝑖, 𝑗) ∗ 𝐵]𝑖, 𝑗 (4)

where 𝑎(𝑖, 𝑗) is the element in the row and 𝑖𝑡ℎ 𝑗 𝑡ℎ
column of matrix 𝐴.
The Kronecker product of two matrices is a much larger matrix, with sizes equal to the product of the sizes of the original
matrices. In this work, authors replaced the large-weight matrices with combinations of multiple Kronecker products of smaller
matrices. Furthermore, the computation efficiency of the Kronecker convolution allows us to utilize larger feature maps for better
classification. If Kronecker and outer products have the same number of picture pixel value parameters, Kronecker will result in less
reconstruction error. For weight matrices and tensors in convolutional networks, our method may generate approximate models that
run faster and have fewer parameters while maintaining the same degree of accuracy loss. The major difference between Kronecker
convolution and standard convolution operation is shown in Fig. 6. In Kronecker convolution, each pixel point on the image is
multiplied by each point on the filter, which creates a unique sub-feature set. In standard convolution, each subset of pixels is
multiplied with the filter, creating a new sample on the feature map as shown in Fig. 6. When the stride value is minimum, the
resulting feature-map size will be high. Large portions of the filters overlapped due to the minimum value of stride; more features
will be shared between the outputs. Consequently, When the stride size is increased, fewer parameters are shared between filters, and
the feature map becomes smaller. Increasing the stride has the effect of down-sampling the image, obscuring lower-level details. In
the proposed technique, extracting deep features without filter overlapping is unable to provide a great advantage in classification.
The designed model combines several layers, including the image input layer, Kronecker convolutional layer (Conv2D), rectifier
linear unit (ReLu) activation layer, max pooling layer, flatten layer, Dropout layer, and Dense layers. Fig. 7 illustrates the proposed
model’s complete architecture. Convolutional neural networks (CNNs) are effective in image classification [32–34]. However, it
requires more data for effective training. In addition to this, CNN can predict the output data with the help of millions of unknown
parameters. As CNN is a block box, it is very hard to understand and can’t handle medical imaging data efficiently. In general,
medical imaging data differs from normal images in the following aspects: (i) fewer images were used for training, and (ii) it is hard
for the designed model to understand when predicting or classifying the output. In this work, a new DKN for kidney stone detection is
suggested to fix the problems listed above with CNN. DKN can accumulate good knowledge with a low sample set of training images,
with good interpretability, and is able to provide better performance than conventional CNN. Most of DKN is based on Kronecker

6
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 7. Proposed DKN architecture for automated kidney stone detection.

products, which indirectly keep coefficients’ latent piece-wise smooth properties. DKN is allowed to find the important parts of the
image so that the model can be understood better. DKN could then manage data from many different types of medical imaging, such
as CT, MRI, fMRI, and others [35]. We call DKN a network because it works like a CNN or a fully convolutional network (FCN).
A Kronecker structure serves as the basis for DKN. However, it is possible to express it in convolutional form as well, and it has
no overlaps between the convolutions, unlike in traditional CNN. This architecture gives us the right model interpretability while
reducing the number of dimensions as much as possible.
The Kronecker product is used instead of a normal convolution operation in this work. In standard deep learning architectures,
overlapped convolutions are utilised when the stride is equal to one. Due to these overlapped convolutions, the same information in
the images repeatedly increased. Therefore, Kronecker product is utilized in this network to remove the redundant information from
the convoluted images. These Kronecker product convolutions are overlapped free as we are multiplying filter with individual pixels
on an image as shown in Fig. 6. The authors reconstructed the L-layer fully convolutional neural network (FCNN) based on these
products. To demonstrate this operation, let us assume A is an image of size 128 × 128 and the convolution filter size of 3 × 3, then
the final convolution operation denoted as
The Image size of 128 × 128 is defined as

⎡ 𝐴1_1 𝐴1_2 ........ 𝐴1_128 ⎤

⎢ ⎥
𝐴 𝐴2_2 ........ 𝐴2_128
𝐴 = ⎢ 2_1 ⎥ (5)
⎢ ........ ........ ........ ........ ⎥
⎢𝐴 𝐴128_128 ⎥
⎣ 128_1 ⎦
The Convolution filter of size 3×3 is defined as

⎡ 𝐶1 𝐶2 𝐶3 ⎤
𝐶 = ⎢ 𝐶4 𝐶5 𝐶6 ⎥ (6)
⎢ ⎥
⎣ 𝐶7 𝐶8 𝐶9 ⎦
Now the convolution operation is operated on every single element of A as like a Kronecker product [36]. The resultant Kronecker
convoluted matrix image of size 128 × 3 × 128 × 3. The detailed diagram related to the above technique is represented in Fig. 8. After
the Kronecker convolution operation on the image in the first step, max-pooling is employed for a down-sampling operation which
reduces the spatial size of the input image by taking the maximum value of a set of pixels in a window. To make the model more
computationally efficient, Max-pooling is employed to reduce the size of the feature maps. Another Kronecker convolution operation
is applied to the output of the max-pooling layer. This helps to extract more complex features from the input image. A dense layer
is a layer in which every neuron is coupled to each neuron in the layer underneath it. This layer aims to map the features extracted
from the Kronecker convolutional layers to a set of output classes. Another max-pooling layer is applied to the output of the dense
layer. This helps to reduce the dimensionality of the input to the next layer. The final Kronecker convolution operation is applied
to the output of the second max-pooling layer. This helps to extract even more complex features from the input image. Another last
max-pooling layer is applied to the output of the third Kronecker convolution layer. This further reduces the dimensionality of the
input. The output of the third max-pooling layer is flattened into a vector. This helps prepare the output for the final classification
layer. This model uses the regularization technique dropout to prevent overfitting. During training, random neurons are eliminated
from the network, forcing the remaining neurons to acquire more robust properties. The final dense layer maps the output of the
previous layers to a set of output classes. This layer uses a softmax activation function to produce class probabilities. The whole
proposed methodology is shown in Algorithm 1 (pseudocode).

7
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 8. A sample of pixel-wise Kronecker convolution.

Table 2
Details of data used in various training, validation, and testing operations.

Data Kidney Stone Normal Total

No. of Subjects 268 165 433

Coronal CT Scans 790 1009 1799
Training and validation 690 900 1590
Testing 100 109 209

Algorithm 1 Kronecker convolutional neural network for kidney stone classification.

1: Collect and pre-process data: Collect a dataset of kidney stone images and pre-process them by resizing them to a standard size and converting them to grayscale.
2: Split the data: Split the dataset into training, validation, and testing sets, typically in an 80-10-10 split, respectively.
3: Define the factors: Define the factor matrices for each layer of the DKN. These factor matrices will be used to define the weights for the horizontal and vertical
convolutions for each layer.
4: Define the model: Define the architecture of the DKN, using the factor matrices defined in step 3. Typically, the network will consist of several Kronecker Conv2D
layers, followed by a flattened layer and one or more dense layers.
5: Design the model: Design the DKN, using appropriate loss, optimizer, and metric functions. The categorical cross-entropy loss function is frequently used for
multi-class classification problems such as the classification of kidney stones.
6: Train the model: Train the DKN using the training set, adjusting the weights of the network to minimize the loss function.
7: Eevaluate the model: Evaluate the performance of the DKN on the validation set, calculating metrics like accuracy, precision, recall, and F1 score.
8: Test the model: Test the DKN on the testing set, reporting metrics like accuracy, precision, recall, and F1 score.
9: Fine-tune the model: Fine-tune the DKN by adjusting the hyperparameters or changing the architecture, based on the results of the evaluation and testing.

4. Results and discussion

We developed the proposed model in the Python programming language, using the PIL library with compact deep learning
libraries Keras, Tensorflow, and open CV. The hardware used in the simulation included an Intel I7 processor, 16 GB of DDR3, and
an NVIDIA RTX 2060 graphics card. There are 1799 coronal CT scans in the database, belonging to 790 kidney stones and 1009
normal healthy individuals. To conduct experiments, we divided the available 1799 CT image data into three different parts training
(80%), validation (10%), and testing (10%). To evaluate the recommended DKN, two different validation strategies, 10-fold cross-
validation and test validation, were performed. In this work, 1590 (80% Train and 10% validation) images were used for training
with cross-validation and 209 (10% test) images for testing. Detailed information about dataset splitting for experimentation is
presented in Table 2. Further 10-fold cross-validation is carried out in order to have a better understanding of the suggested model’s
capacity for prediction when it is supplied with new data. In addition, the process of tuning hyperparameters involves iterating
through a wide variety of hyperparameter combinations and then choosing the range of values for each parameter for which k-folds
generate the most accurate results. Table 3 contains a listing of the performance indicators that were obtained from the 10-fold
cross-validation, along with their mean values. During the validation set, many folds performed admirably, including Fold-2, Fold-4,
and Fold-6.

4.1. Optimization of hyperparameters

Hyperparameter tuning is performed during the validation phase of a deep learning model’s training. After training a model on
the training dataset, the validation dataset is used to assess its performance and tune the model’s hyperparameters to optimize its
performance. This involves adjusting the hyperparameters, such as the learning rate (𝜂), batch size, and the number of epochs, and
evaluating the model’s performance on the validation set. This process is repeated until the best combination of hyperparameters are
found, and the model’s performance on the test set is used as a final evaluation metric to ensure that the model is not overfitting

8
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Table 3
Performance parameters using 10-fold cross validation.

Parameter Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Fold 6 Fold 7 Fold 8 Fold 9 Fold 10 Avg.±SD

Accuracy (%) 99.52 100 99.14 100 98.95 100 98.75 99.59 99.18 99.19 99.432±0.460
Precision (%) 99.58 99.88 100 99.16 99.24 98.89 100 99.54 99.55 99.72 99.556±0.369
Sensitivity (%) 99.61 100 99.45 98.92 100 98.94 99.39 99.48 99.81 98.84 99.444±0.431
Specificity (%) 100 99.48 99.38 99.48 100 99.35 98.68 99.31 99.14 99.32 99.414±0.385
F1-score (%) 99.72 100 99.29 99.42 98.82 99.22 99.84 99.72 98.82 99.61 99.446±0.408

Table 4
Performance of the model for various epochs.

No. of epochs Learning rate Batch size Accuracy (%)

20 0.001 16 85.71
40 0.001 16 91.19
70 0.001 16 95.44
100 0.001 16 97.92
120 0.001 16 99.43
130 0.001 16 98.88

Table 5
Performance of the model obtained for various different learning rates.

No. of epochs Learning rate Batch size Accuracy (%)

120 0.1 16 96.82

120 0.2 16 95.19
120 0.05 16 97.92
120 0.01 16 98.45
120 0.005 16 98.91
120 0.001 16 99.43

Table 6
Performance of the model obtained for different batch sizes.

No. of epochs Learning rate Batch size Accuracy (%)

120 0.001 256 97.12

120 0.001 128 97.55
120 0.001 64 98.14
120 0.001 32 98.92
120 0.001 16 99.43
120 0.001 8 98.29

to the validation set. Initially, the network was evaluated by varying the number of epochs while maintaining a constant batch size
and learning rate. As shown in Table 4, the performance of the proposed architecture reached 99.43% with 120 epochs. The system
was then tested with different 𝜂 while keeping the batch size and a constant number of epochs. Table 5 shows that the optimal 𝜂 is
0.001. Similarly, the batch size of the network was varied while holding the other two parameters constant, and Table 6 shows that
the proposed model’s optimal batch size is 16.
Improved classification performance was observed during the training phase through a series of experiments by employing the
root mean squared propagation (RMSprop) training approach with the following optimized hyperparameters: 𝜂 of 0.001, 120 epochs,
and a batch size of 16. The proposed architecture’s detailed parameters are shown in Table 7. Fig. 9 presents a graph of the training
accuracy and validation accuracy for each training epoch for the proposed CNN model. The associated confusion matrices 10-fold
cross-validation (training) and testing are shown in Fig. 10.
The confusion matrix is a two-dimensional matrix where each column represents an actual class, and each row represents a
prediction class [5,37,38]. Table 8 provides complete information regarding the relevant performance parameters for (a) 10-fold CV
and testing. The findings of the suggested CNN for kidney stone detection reflect considerable performance by achieving accuracies
of 99.43% and 98.55% by adopting 10-fold CV and testing validation techniques, respectively. This indicates that the proposed
CNN is capable of producing accurate results. Fig. 11 illustrates the relevant receiver operating characteristic (ROC) curve and the
precision-recall (PR) curve for classification. The use of a precision-recall curve is recommended for datasets with an unbalanced
distribution, whereas the use of ROC curves is recommended for datasets with a balanced distribution. In this work, the authors used
an unbalanced dataset (790 kidney stone scans and 1009 normal scans) to validate the performance of the developed model. The
correctly classified and misclassified kidney stone images are shown in Fig. 12. It may be noted from this figure that, the proposed
method is able to classify the kidney stone image perfectly when the input image has good quality. In Fig. 12, (a), (b), (c), and (d)

9
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Table 7
Hyperparameters used for this work.

Hyperparameters of our Kronecker network Details

epochs number 120

Batch size 16
Learning rate 0.001
Stride 1
Dilation rate 1
CV technique 10-fold
Number of kernels in layer one 64
Activation ReLu
Dropout rate 0.5
Number of trainable parameters 2,71,152
Number of non-trainable parameters 58,652
Average Training time 3122 Seconds

Fig. 9. Training and validation performance curves of the proposed technique for kidney stone detection.

Table 8
Performance metrics obtained for the proposed DKN kidney stone detection system.

Performance measures Training and CV (10-fold) Testing

Accuracy (%) 99.43±0.460 98.56

Precision (%) 99.56±0.369 99.07
Sensitivity (%) 99.45±0.431 98.15
Specificity (%) 99.42±0.385 98.99
F1-score (%) 99.50±0.408 98.60
False Positive Rate 0.0058 0.0101
False Negative Rate 0.0055 0.0185
Matthews Correlation Coefficient 0.9885 0.9710

images have correctly identified the kidney stones, and it is misclassified (e) and (f) kidney stone images due to their poor quality.
The quality of the image plays a significant role in classification. The limitations of this study are given below

1. The network may struggle to classify them if the input images are blurry or have low contrast.
2. The lack of diversity in the training data. Training data consists of images of kidney stones that are all similar in appearance.
Hence the network may not be able to distinguish between different types of kidney stones.
3. Kidney stones can come in various shapes, sizes, and types. If the network has not been trained on a diverse range of kidney
stone images, it may struggle to classify certain types of stones that it has not seen before.

In future, we are planning to extend this work using more images collected from different centers belonging to various races.
Also, we plan to explore the possibility of providing trust in our diagnosis by implementing explainable artificial intelligence (XAI)
[39]. This will help to visualize the kidney stone location clearly to the clinicians. Our deep learning model is prone to noise, and
hence the tuning parameters need to be tuned to obtain high performance in real-world scenarios. Uncertainty quantification (UQ)
can be employed to quantify and overcome the uncertainty in the deep learning models [40].

10
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 10. Confusion Matrices generated using the proposed CNN model with (a) 10-fold cross-validation (Training) and (b) Testing.

Table 9
Performance Comparison of the traditional approaches with the proposed approach.

Network Accuracy (in %) Sensitivity (in %) Specificity (in %) Precision (in %) F1-Score (in %) Epochs Testing time for one image

CNN 93.22 92.14 89.90 91.22 90.43 200 1.202 Sec

ResNet-50 94.87 93.44 91.56 93.38 92.97 180 0.987 Sec
Proposed DKN 98.56 98.15 98.99 99.07 98.60 150 1.587 Sec

4.2. Computational complexity

The computational complexity in deep learning models for image recognition with large datasets or complex architectures will be
intensive. The Kronecker convolutions used in this paper can also be more computationally demanding than traditional convolutions
due to the use of Kronecker products. The computational overhead is related to the number of parameters, and training time is
shown in Table 7. Therefore, it is important to consider the hardware and software resources required to develop the model. Also,
employ optimization techniques to reduce the computing overhead. Specialized hardware such as GPUs or TPUs, implementing
parallelization techniques, or reducing the complexity of the architecture through techniques such as pruning or compression can be
employed.

4.3. A comparative analysis of the proposed approach with standard deep learning architectures

In this section, the proposed approach is compared with the standard CNN and ResNet. Generally, CNN provides better perfor-
mance for image classification even though sometimes we need more depth to get an effective classification. Increasing the depth
of the network is not the only solution, and it will create new problems, such as vanishing gradient. Therefore, a new architecture
such as ResNet is introduced to manage this gradient vanishing problem efficiently. However, ResNet increases the computational
complexity and cost due to skipping connections. Hence, the authors introduced DKN based on the Kronecker product to solve the
problems of traditional architecture. We designed DKN with three convolution layers, three max-pooling, one flattened, three dense,
and one dropout layer. Here all the convolutions are Kronecker product-based convolutions. The authors are experimented with CNN
with the same number of layers by using ordinary convolutions, and ResNet also designed with skip connections. The results of these
experiments are summarized in Table 9 in terms of performance metrics, computational time and complexity.
It can be observed from Table 9 it is observed that the proposed method performed better than the standard deep learning
architectures. However, the proposed DKN is taking a longer time to identify the test image compared to the standard architectures,
even though the accuracy of the DKN is very high compared to them. This is the major limitation of the proposed network, even

11
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 11. Classification performance curves: (a) ROC and (b) Precision-Recall of the proposed DKN.

though an exact diagnosis is important compared to the time complexity of biomedical applications. Hence, the proposed end-to-
end system is able to detect kidney stones in coronal CT images accurately and assist the radiologists in their diagnosis. The major
advantages of the KNN are described as follows.

• Computational efficiency: Kronecker convolution can be implemented using fast matrix multiplication, which makes it faster
and more efficient than traditional convolution.
• Sparsity: Kronecker convolution is particularly well-suited to sparse data, as it is able to take advantage of the sparsity of the
input data and reduce the number of computations required.
• Memory efficiency: Kronecker convolution can also be more memory-efficient than traditional convolution, as it allows for
weight sharing and parameter sharing between different layers of the CNN.
• Model interpretability: Kronecker convolution can also make it easier to interpret the learned features of a CNN, as it allows for
the explicit separation of local and global features in the model.
• Generalization: Kronecker convolution has been shown to have good generalization properties, meaning it can perform well on
new, unseen data.

The deep learning architecture utilized in this work involves the application of Kronecker convolutions for the automated detection
of kidney stones with coronal CT images. Kronecker convolutions are a type of CNN that uses Kronecker products to extract more
complex and informative features from the images. The architecture includes multiple layers of Kronecker convolutions, followed by
activation functions and pooling layers to further refine the extracted features. The main advantages of Kronecker convolution are
its computational efficiency, sparsity, memory efficiency, interpretability, and generalization.

4.4. Comparison of the proposed model with existing similar literature

In most cases, medical image processing and diagnosis are performed with the assistance of convolution neural networks [41].
In the past, authors have been utilised in a variety of medical imaging modalities, including MRI, CT, and X-ray, for brain tumour

12
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Fig. 12. Kidney stone image classification: (a), (b), (c), and (d) are the final classified and (e) and (f) are the misclassified kidney stone CXR images by the proposed
methodology.

Table 10
Overview of the comparison with similar works.

Parameter Yildirim et al. [15] Baygin et al. [16] Proposed method*

Image modality Coronal CT scans Coronal CT scans Coronal CT scans

Dataset 1799 CT Images (790 Kidney stone 1009 1799 CT Images (790 Kidney stone 1009 1799 CT Images (790 Kidney stone 1009
Normal healthy) Normal healthy) Normal healthy)

Image Processing Not considered Resized CT image to 224 x 224 Distinct image enhancing techniques such
methods as image embossing, region identification,
and cropping were developed

Data Augmentation Yes (rotation 10 degrees zooming) Not considered Yes (rotation 30 degrees,
flipping-horizontal shear- 0.3 and
zoom-0.3)

Data Splitting 80% training 20% testing 80% training 20% testing 80% training, 10% validation 10% testing

Type of Deep CNN XResNet-50 (Transfer Learning) ExDark19 model (Transfer Learning based Customized deep CNN (Based on a novel
model image classification model) Kronecker product structure)

Training Parameters Epochs-40 Iterations-30 Epochs-150

Optimizer-ADAM Optimizer-Bayesian Optimizer-RMS prop
(Other hyperparameters details are not Learning rate-0.001
reported) Batch Size-16

Performance metrics Accuracy-96.82% Accuracy-99% Accuracy-98.56%

Sensitivity-95.76% Sensitivity-98.35% Sensitivity-98.15%
Specificity-97.79% Specificity-99.90% Specificity-98.99%
Precision-97.53% Precision-99.87% Precision-99.07%
F1-score-96.44% F1-score-99.10% F1-score-98.60%
MCC-0.93 MCC-0.97

Remarks System complexity is high System complexity is high and Simple intelligent classification system
time-consuming (Less complex) with average training time
is 3122 sec.

segmentation [42], prostate cancer [43], pancreatic segmentation [44], breast cancer [45], and liver disease segmentation [46]. Only
a few authors have employed deep learning on coronal CT scans to diagnose kidney stones.
Data acquisition, pre-processing for the region of interest, and machine learning or deep model classification are the standard
steps in most research on the computer-aided detection of kidney stones. A deep convolutional neural network, such as the ResNet-
101 package, was used in work by Black KM et al. [17] to detect kidney stones in digital images automatically. A computer vision

13
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

Table 11
Ablation study of the proposed DKN.

Normal Kronecker Normal Kronecker Normal Kronecker Acc Sen Spe Pre F1-score
convolution convolution convolution convolution convolution convolution (%) (%) (%) (%)
Layer-1 Layer-1 Layer-2 Layer-2 Layer-3 Layer-3
✓ × ✓ × ✓ × 93.22 92.14 89.90 91.22 0.904
× ✓ ✓ × ✓ × 95.12 94.23 93.86 95.01 0.939
× ✓ × ✓ ✓ × 97.91 96.87 95.97 96.82 0.969
× ✓ × ✓ × ✓ 98.56 98.15 98.99 99.07 0.986

technique was proposed by Thein et al. [47] to find kidney stones in CT scans and separate them. A sensitivity of 95.24% was reported
after applying three thresholding algorithms to eliminate undesired zones based on intensity, size, and location.
Akshaya et al. [48] used a machine learning approach, namely a back propagation neural network, to identify kidney stones.
Feature extraction and image classification were used stages in the decision-making process. The principal component analysis was
used for feature extraction, while the backpropagation neural network (BPN) was used for classification. In addition, a segmentation
strategy based on the Fuzzy C-Mean (FCM) clustering algorithm was used to detect kidney stones from CT images. Several methods,
including the use of imaging modalities, were proposed by Manoj et al. [49]. They used VGG16 to automatically detect kidney stones
from CT images. Moreover, the stratified K-fold cross-validation technique has been used to develop the model. Grad-CAM, short for
gradient-weighted class activation mapping, is then used to pinpoint the precise location of the kidney stone [50]. The similar works
[15,16] are compared with the proposed work in Table 10. From Table 10, it is observed that the proposed method performs better
compared to similar existing works.
In the present work, a DKN is proposed to detect kidney stones in coronal CT images. The main contribution of the work is to
approach the automatic detection of kidney stones by using a comprehensive learning algorithm by utilizing significant segmented
image features from low-resolution CT scan data. All convolutions in the proposed deep CNN model are Kronecker convolutions,
which are based on a novel Kronecker product structure. The proposed method achieved a remarkable 98.56% accuracy on 209
test CT images. Also, we compared our proposed method to a very similar study that was published by authors in [15]. The major
advantage of the proposed method is that it offers large, in-depth feature maps, which improve the performance of the proposed DKN.
When Kronecker convolution is used on every pixel, it makes a unique sub-feature map. This lets the network classify accurately.
The only disadvantage of the proposed method is the number of mathematical manipulations. When compared to the traditional
CNN, it takes almost twice as many calculations to make effective feature maps from the input data. In the future, authors will try to
decrease the computational complexity of the Kronecker convolution.

5. Ablation study of the proposed DKN to identify the importance of Kronecker convolution

An ablation study is a technique used to evaluate the importance of individual components or factors in a complex system. In the
context of deep learning, an “ablation study” is a type of experiment where the performance of a deep learning model is evaluated
by selectively removing or “ablat(ing)” certain components or features from the model. The main significance of the ablation study
in this work is as follows.

• To understand the impact of individual components or features on the overall performance of the model.
• To identify the most important components or features for a particular task or dataset.
• To determine the minimum number of components or features needed to achieve a certain level of performance.
• To guide the design of future models by identifying which components or features are most important.

A proposed network is experimented with using a baseline model, and then certain components or features are systematically
removed or modified to see how the model’s performance is affected. This can be done by training and evaluating multiple versions
of the model with different combinations of components or features. Overall, ablation studies are a good way to learn how different
parts or features of a deep learning model affect its performance. They can also help guide the design of future models.
Table 11 presents a summary of the ablation investigation and its impact on the suggested DKN’s performance. Research showed
that incorporating a Kronecker convolution layer into the network’s later stages improved its overall effectiveness. Including three
Kronecker convolution layers improves the CNN network’s performance from around 93.22% with normal convolution layers to
98.56%.

6. Conclusion

This work proposes a novel Kronecker product structure-based custom CNN model for automatic kidney stone diagnosis. In this
work, the performance of the proposed DKN is compared with traditional approaches like CNN, ResNet, and AlexNet in this work. The
proposed architecture is able to perform accurately without increasing the depth or skipping connections, as in residual and deeper
networks. Kronecker convolutions effectively extract useful features from the input images, which are the best source of model
training. The experimentation was carried out using 1799 coronal CT scans in the database (790 kidney stones and 1009 normal
healthy) collected from 433 individuals. The proposed method was validated using a 10-fold CV and attained a testing accuracy of

14
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

98.56%. The developed model is a clinically useful diagnostic tool since it has achieved maximum classification performance. As a
result, our proposed deep CNN model is effective and can help radiologists to identify kidney stone cases accurately.

CRediT authorship contribution statement

1. Conceptualization Methodology, Software Programming: AJP, KKP, PP, 2. Validation, Verification: UCRA, AJP, KKP Formal
analysis: BCN, KKP, MH 3. Investigation: AJP, KKP, UCRA, OY 4. Resources: AJP, PP 5. Writing - Original Draft Preparation: AJP,
KKP 6. Writing - Review & Editing: AJP, KKP, UCRA, BCN 7. Visualization & Supervision: AJP, PP, RT 8. Project administration: AJP,
PP.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgments

Kiran Kumar Patro would like to thank Aditya Institute of Technology and Management, Tekkali, Srikakulam for their support,
and Mohamed Hammad would like to thank Prince Sultan University for their support.

References

[1] I. Sorokin, C. Mamoulakis, K. Miyazawa, A. Rodgers, J. Talati, Y. Lotan, Epidemiology of stone disease across the world, World J. Urol. 35 (9) (2017) 1301–1320.
[2] T. Alelign, B. Petros, Kidney stone disease: an update on current concepts, Adv. Urol. 2018 (2018).
[3] N. Thein, H.A. Nugroho, T.B. Adji, K. Hamamoto, An image preprocessing method for kidney stone segmentation in ct scan images, in: 2018 International
Conference on Computer Engineering, Network and Intelligent Multimedia (CENIM), IEEE, 2018, pp. 147–150.
[4] K.R. Pedada, B. Rao, K.K. Patro, J.P. Allam, M.M. Jamjoom, N.A. Samee, A novel approach for brain tumour detection using deep learning based technique,
Biomed. Signal Process. Control 82 (2023) 104549.
[5] K.K. Patro, A. Jaya Prakash, M. Jayamanmadha Rao, P. Rajesh Kumar, An efficient optimized feature selection with machine learning approach for ecg biometric
recognition, IETE J. Res. (2020) 1–12.
[6] R. Kijowski, F. Liu, F. Caliva, V. Pedoia, Deep learning for lesion detection, progression, and prediction of musculoskeletal disease, J. Magn. Reson. Imaging
52 (6) (2020) 1607–1619.
[7] A.K. Poyraz, S. Dogan, E. Akbal, T. Tuncer, Automated brain disease classification using exemplar deep features, Biomed. Signal Process. Control 73 (2022)
103448.
[8] V. Ravi, Attention cost-sensitive deep learning-based approach for skin cancer detection and classification, Cancers 14 (23) (2022) 5872.
[9] V. Ravi, H. Narasimhan, T.D. Pham, A cost-sensitive deep learning-based meta-classifier for pediatric pneumonia classification using chest x-rays, Expert Syst.
39 (7) (2022) e12966.
[10] V. Ravi, V. Acharya, M. Alazab, A multichannel efficientnet deep learning-based stacking ensemble approach for lung disease detection using chest x-ray images,
Clust. Comput. (2022) 1–23.
[11] E. Skounakis, K. Banitsas, A. Badii, S. Tzoulakis, E. Maravelakis, A. Konstantaras, Atd: a multiplatform for semiautomatic 3-d detection of kidneys and their
pathology in real time, IEEE Trans. Human-Mach. Syst. 44 (1) (2013) 146–153.
[12] S. Yin, Q. Peng, H. Li, Z. Zhang, Xinge You, Katherine Fischer, Susan L. Furth, Gregory E. Tasian, Yong Fan, Automatic kidney segmentation in ultrasound images
using subsequent boundary distance regression and pixelwise classification networks, Med. Image Anal. 60 (2020) 101602.
[13] V.B. Kolachalama, P. Singh, C.Q. Lin, D. Mun, M.E. Belghasem, J.M. Henderson, J.M. Francis, D.J. Salant, V.C. Chitalia, Association of pathological fibrosis with
renal survival using deep neural networks, Kidney Int. Rep. 3 (2) (2018) 464–475.
[14] N.A. Almansour, H.F. Syed, N.R. Khayat, R.K. Altheeb, R.E. Juri, J. Alhiyafi, S. Alrashed, S.O. Olatunji, Neural network and support vector machine for the
prediction of chronic kidney disease: a comparative study, Comput. Biol. Med. 109 (2019) 101–111.
[15] K. Yildirim, P.G. Bozdag, M. Talo, O. Yildirim, M. Karabatak, U.R. Acharya, Deep learning model for automated kidney stone detection using coronal ct images,
Comput. Biol. Med. 135 (2021) 104569.
[16] M. Baygin, O. Yaman, P.D. Barua, S. Dogan, T. Tuncer, U.R. Acharya, Exemplar darknet19 feature generation technique for automated kidney stone detection
with coronal ct images, Artif. Intell. Med. 127 (2022) 102274.
[17] K.M. Black, H. Law, A. Aldoukhi, J. Deng, K.R. Ghani, Deep learning computer vision algorithm for detecting kidney stone composition, BJU Int. 125 (6) (2020)
920–924.
[18] M. Längkvist, J. Jendeberg, P. Thunberg, A. Loutfi, M. Lidén, Computer aided detection of ureteral stones in thin slice computed tomography volumes using
convolutional neural networks, Comput. Biol. Med. 97 (2018) 153–160.
[19] L.A. Fitri, F. Haryanto, H. Arimura, C. YunHao, K. Ninomiya, R. Nakano, M. Haekal, Y. Warty, U. Fauzi, Automated classification of urinary stones based on
microcomputed tomography images using convolutional neural network, Phys. Med. 78 (2020) 201–208.
[20] H. Xiang, Q. Chen, Y. Wu, D. Xu, S. Qi, J. Mei, Q. Li, X. Liu, Urine calcium oxalate crystallization recognition method based on deep learning, in: 2019
International Conference on Automation, Computational and Technology Management (ICACTM), IEEE, 2019, pp. 30–33.
[21] Y. Kazemi, S.A. Mirroshandel, A novel method for predicting kidney stone type using ensemble learning, Artif. Intell. Med. 84 (2018) 117–126.
[22] S. Chaitanya, P. Rajesh Kumar, Detection of chronic kidney disease by using artificial neural networks and gravitational search algorithm, in: Innovations in
Electronics and Communication Engineering, Springer, 2019, pp. 441–448.
[23] M. Suresh, M. Abhishek, Kidney stone detection using digital image processing techniques, in: 2021 Third International Conference on Inventive Research in
Computing Applications (ICIRCA), IEEE, 2021, pp. 556–561.

15
K.K. Patro, J.P. Allam, B.C. Neelapu et al. Information Sciences 640 (2023) 119005

[24] M. Kahani, S.H. Tabrizi, A. Kamali-Asl, S. Hashemi, A novel approach to classify urinary stones using dual-energy kidney, ureter and bladder (dekub) x-ray
imaging, Appl. Radiat. Isot. 164 (2020) 109267.
[25] Y.-Y. Liu, Z.-H. Huang, K.-W. Huang, Deep learning model for computer-aided diagnosis of urolithiasis detection from kidney–ureter–bladder images, Bioengi-
neering 9 (12) (2022) 811.
[26] M. Kobayashi, J. Ishioka, Y. Matsuoka, Y. Fukuda, Y. Kohno, K. Kawano, S. Morimoto, R. Muta, M. Fujiwara, N. Kawamura, et al., Computer-aided diagnosis
with a convolutional neural network algorithm for automated detection of urinary tract stones on plain x-ray, BMC Neurol. 21 (1) (2021) 1–10.
[27] W. Brisbane, M.R. Bailey, M.D. Sorensen, An overview of kidney stone imaging techniques, Nat. Rev. Urol. 13 (11) (2016) 654–662.
[28] S. Nikneshan, S. Mohseni, M. Nouri, H. Hadian, M.J. Kharazifard, The effect of emboss enhancement on reliability of landmark identification in digital lateral
cephalometric images, Iran. J. Radiol. 12 (2) (2015).
[29] M. Kim, J. Yun, Y. Cho, K. Shin, R. Jang, H.-j. Bae, N. Kim, Deep learning in medical imaging, Neurospine 16 (4) (2019) 657.
[30] C.F. Van Loan, The ubiquitous kronecker product, J. Comput. Appl. Math. 123 (1–2) (2000) 85–100.
[31] T. Wu, S. Tang, R. Zhang, J. Cao, J. Li, Tree-structured kronecker convolutional network for semantic segmentation, in: 2019 IEEE International Conference on
Multimedia and Expo (ICME), IEEE, 2019, pp. 940–945.
[32] A.J. Prakash, Capsule network for the identification of individuals using quantized ecg signal images, IEEE Sens. Lett. 6 (8) (2022) 1–4.
[33] Allam Jaya Prakash, K.K. Patro, Samantray Saunak, Pradipta Sasmal, P Lalitha Kumari, Tummalapalli Geetamma, A new approach of transparent and explainable
artificial intelligence technique for patient-specific ECG beat classification, IEEE Sens. Lett. 7 (5) (2023) 1–4.
[34] K.K. Patro, A.J. Prakash, M. Hammad, R. Tadeusiewicz, P. Pławiak, Scovnet: a skip connection-based feature union deep learning technique with statistical
approach analysis for the detection of covid-19, Biocybern. Biomed. Eng. (2023).
[35] M.J. Ali, B. Raza, A.R. Shahid, Multi-level kronecker convolutional neural network (ml-kcnn) for glioma segmentation from multi-modal mri volumetric data,
J. Digit. Imag. 34 (4) (2021) 905–921.
[36] Y. Zhang, Y. Li, S. Wang, C.-X. Wang, Image restoration based on the convolution kernel matrix reconstructed by kronecker product, in: 2020 IEEE 5th
International Conference on Signal and Image Processing (ICSIP), IEEE, 2020, pp. 185–189.
[37] A.J. Prakash, K.K. Patro, M. Hammad, R. Tadeusiewicz, P. Pławiak, Baed: a secured biometric authentication system using ecg signal based on deep learning
techniques, Biocybern. Biomed. Eng. 42 (4) (2022) 1081–1093.
[38] K.K. Patro, A.J. Prakash, S. Samantray, J. Pławiak, R. Tadeusiewicz, P. Pławiak, A hybrid approach of a deep learning technique for real–time ecg beat detection,
Int. J. Appl. Math. Comput. Sci. 32 (3) (2022) 455–465.
[39] H.W. Loh, C.P. Ooi, S. Seoni, P.D. Barua, F. Molinari, U.R. Acharya, Application of explainable artificial intelligence for healthcare: a systematic review of the
last decade (2011–2022), Comput. Methods Programs Biomed. (2022) 107161.
[40] V. Jahmunah, E. Ng, R.-S. Tan, S.L. Oh, U.R. Acharya, Uncertainty quantification in densenet model using myocardial infarction ecg signals, Comput. Methods
Programs Biomed. 229 (2023) 107308.
[41] J. Wang, H. Zhu, S.-H. Wang, Y.-D. Zhang, A review of deep learning on medical image analysis, Mob. Netw. Appl. 26 (1) (2021) 351–380.
[42] R. Ranjbarzadeh, A. Bagherian Kasgari, S. Jafarzadeh Ghoushchi, S. Anari, M. Naseri, M. Bendechache, Brain tumor segmentation based on deep learning and
an attention mechanism using mri multi-modalities brain images, Sci. Rep. 11 (1) (2021) 1–17.
[43] T. Kiljunen, S. Akram, J. Niemelä, E. Löyttyniemi, J. Seppälä, J. Heikkilä, K. Vuolukka, O.-S. Kääriäinen, V.-P. Heikkilä, K. Lehtiö, et al., A deep learning-based
automated ct segmentation of prostate cancer anatomy for radiation therapy planning-a retrospective multicenter study, Diagnostics 10 (11) (2020) 959.
[44] Y. Zhang, J. Wu, Y. Liu, Y. Chen, W. Chen, E.X. Wu, C. Li, X. Tang, A deep learning framework for pancreas segmentation with multi-atlas registration and 3d
level-set, Med. Image Anal. 68 (2021) 101884.
[45] E.H. Houssein, M.M. Emam, A.A. Ali, P.N. Suganthan, Deep and machine learning techniques for medical imaging-based breast cancer: a comprehensive review,
Expert Syst. Appl. 167 (2021) 114161.
[46] Q. Yan, B. Wang, W. Zhang, C. Luo, W. Xu, Z. Xu, Y. Zhang, Q. Shi, L. Zhang, Z. You, Attention-guided deep neural network with multi-scale feature fusion for
liver vessel segmentation, IEEE J. Biomed. Health Inform. 25 (7) (2020) 2629–2642.
[47] N. Thein, H.A. Nugroho, T.B. Adji, K. Hamamoto, An image preprocessing method for kidney stone segmentation in ct scan images, in: 2018 International
Conference on Computer Engineering, Network and Intelligent Multimedia (CENIM), IEEE, 2018, pp. 147–150.
[48] M. Akshaya, R. Nithushaa, N.S.M. Raja, S. Padmapriya, Kidney stone detection using neural networks, in: 2020 International Conference on System, Computation,
Automation and Networking (ICSCAN), IEEE, 2020, pp. 1–4.
[49] B. Manoj, N. Mohan, S. Kumar, et al., Automated detection of kidney stone using deep learning models, in: 2022 2nd International Conference on Intelligent
Technologies (CONIT), IEEE, 2022, pp. 1–5.
[50] H.W. Loh, C.P. Ooi, S. Seoni, P.D. Barua, F. Molinari, U.R. Acharya, Application of explainable artificial intelligence for healthcare: a systematic review of the
last decade (2011–2022), Comput. Methods Programs Biomed. (2022) 107161.

16