Multimedia Tools and Applications

https://doi.org/10.1007/s11042-024-19113-y

Automatic lung cancer detection using hybrid particle snake

swarm optimization with optimized mask RCNN

R. Sudha1 · K. M. Uma Maheswari2

Received: 20 July 2023 / Revised: 28 January 2024 / Accepted: 27 March 2024

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024

Abstract
As a result of its aggressive nature and late identification at advanced stages, lung cancer
is one of the leading causes of cancer-related deaths. Lung cancer early diagnosis is a
serious and difficult challenge that is crucial to a person’s survival. The first diagnosis
of the malignant nodules is typically made using chest radiography (X-rays) and com-
puted tomography (CT) scans; however, the potential presence of benign nodules results
in incorrect conclusions. The early phases of both benign and malignant nodules exhibit
striking similarities. In this paper, a novel deep learning-based model is proposed for the
precise diagnosis of malignant nodules. The proposed approach consists of two stages
namely, pre-processing and lung nodule detection. Initially, the Lung CT scan images
are collected from the dataset. Then, to remove the noise present in the input image, we
apply an adaptive median filter. Then, to enhance the image, Contrast Limited Adaptive
Histogram Equalization (CLAHE) is applied. After pre-processing, the image is given
to the optimized mask RCNN classifier to detect the malignant and benign nodules. To
enhance the performance of the Mask RCNN classifier, the hyper-parameters are opti-
mally selected using hybrid particle snake swarm optimization (­PS2OA). The proposed
­PS2OA is a hybridization of particle swarm optimization (PSO) and snake swarm optimi-
zation (SSO). The performance of the proposed approach is analyzed based on different
metrics and effectiveness compared with state-of-the-art works. The proposed approach
attained the maximum accuracy of 97.67%. This work aimed at assisting radiologists to
detect and diagnose small-size pulmonary nodules more accurately.

Keywords Lung cancer · Benign nodules · Malignant nodules · Optimized mask RCNN ·
Snake swarm optimization · Particle swarm optimization

* K. M. Uma Maheswari
[email protected]
R. Sudha
[email protected]
1
Department of Computer Science and Engineering, School of computing, College of engineering
and technology, SRM Institute of Science and Technology, Kattankulathur, India
2
Department of Computing Technologies, School of Computing, College of engineering
and technology, SRM Institute of Science and Technology, Kattankulathur, India

13
Vol.:(0123456789)
 Multimedia Tools and Applications

1 Introduction

According to global cancer statistics for 2020, Lung cancer has been the most common and
lethal oncological illness in the world for many years. According to clinical research, if late-
stage patients had been diagnosed and treated earlier, their survival rate within five years
would climb to 52% from the current 10% to 16% range. Lung nodules are a key indica-
tor of early-stage lung cancer, which can be evident on CT scans as localized, spherical
shaped like lung shadows, and the size is no larger than 3 cm wide [1]. Notwithstanding,
the nano size of lung nodules, their morphology, brightness, and other features are close to
those of the vascular system and other tissues in the pulmonary parenchyma; therefore, phy-
sicians must carefully examine and screen each nodule individually; this procedure is cum-
bersome and easily leads to exhaustion, thereby increasing the chances of diagnostic errors.
Therefore, it is vital to build an automatic detection method to assist physicians in improving
the performance and accuracy of lung nodule diagnosis [2]. On grounds of these details, the
movement in medicine toward Computer-Aided Diagnosis (CAD) systems, which are sub-
ject to quantitative analysis of CT lung images, can improve Lung CT image understanding,
disease diagnosis, and detection of small malignant nodules (which are difficult for a clini-
cian to notice), and diagnostic time [3]. The latest generation of CAD systems also helps in
the screening process to detect Lung Nodules differentiating between benign and malignant
nodules. CAD uses the Artificial Intelligence (AI) algorithm of Deep Learning methods that
efficiently leverage object detection. Deep learning is a robust technique in machine learn-
ing in which the object detector automatically learns the image characteristics required for
computer vision tasks. There are several available algorithms for object detection using deep
learning, including Faster R-CNN, you only look once (YOLO), and single shot detection
(SSD). Mask R-CNN is an improved version of Faster R-CNN [4] and it simultaneously
generates a high-quality segmentation mask for each detected object in an image. It incorpo-
rates the two-stage object detection techniques, in the first stage, RoI is predicted using the
Region Proposal Network (RPN), and in the second stage, the class and box offset values are
predicted in parallel, also producing a segmentation mask for each RoI.
Any CNN model’s performance is impacted by various factors, including the size of the
dataset, the number of classes, the model’s weights, hypermeters, the optimizer, and many oth-
ers. Optimizing hyper-parameter plays a vital role during the training of Convolutional neural
networks [5]. To what extent a convolutional neural network performs well relies on its archi-
tecture and the values of its hyper-parameters. CNN includes many hyper-parameters, based
on the structure and training such as the number of convolution layers, the number of filters,
the size of each filter, Batch size, Learning Rate, momentum, etc. [6, 7]. As not in the model
parameter, hyper-parameter tuning can be done manually but this is a tedious and time-con-
suming process. We can use the automated tuning method to optimize the hyper-parameter
[8] to overcome this. Nowadays Modern optimization techniques, namely the heuristic and
metaheuristic algorithms are applied for optimizing objective functions. However, the heuris-
tic methods have numerical inefficiency in the search process, like high dimension problems,
which leads the complicity in the model [9]. To address this, meta-heuristics and Swarm Intel-
ligence (SI) methodologies and variants were proposed to handle a variety of flexible real-
world optimization tasks and address complex/large-scale optimization issues [10].
The key benefit of the SI optimization algorithms over the deterministic approach is the
randomization introduced throughout the search phase can get stuck in circumstances with
no global ideal solution. Therefore, obtaining the global best solution is practically signifi-
cant in SI [11]. The Particle Swarm Optimization (PSO) algorithm is one of the variants of

13
Multimedia Tools and Applications

SI proposed by James Kennedy, It is used in many real-world applications like integrating

optimization of nonlinear functions and training of neural networks [12] The idea of flying
potential solutions across hyperspace while speeding toward "better" solutions is exclusive
to the concept of particle swarm optimization. In PSO, to achieve the optimal solution, it
uses the global best it’s possible to become stuck in the local optima [13], which has the
problem of premature convergence and doesn’t provide diversity, to avoid this we can use a
trajectory-based technique called snake swarm optimization to avoid in stuck with the local
optima. Snake swarm optimization works based on the behavior of snakes, which can be
incorporated with MRCNN to attain a certain level of accuracy with a significant speedup
in computing [14, 15]. So hybrid P ­ S2OA optimization technique can better seek the global
optimal solution and successfully prevent particles from remaining at the local optimum
[16]. The main contribution of the proposed approach is listed below;

• Initially, the Lung CT scan images are collected from the dataset and to remove the
noise present in the input image, we apply an adaptive median filter. To enhance the
image quality, CLAHE is applied.
• After pre-processing, the image is given to the optimized mask RCNN classifier to
detect the malignant and benign nodules.
• To enhance the performance of the Mask RCNN classifier, the hyper-parameters are
optimally selected using hybrid particle snake swarm optimization algorithm ­(PS2OA).
The proposed ­PS2OA is a hybridization of particle swarm optimization (PSO) and
snake swarm optimization (SSO).
• The performance of the proposed approach is analyzed based on different metrics and
effectiveness compared with state-of-the-art works.

The rest of the paper is structured as follows: Sect. 2 discusses related work; Sect. 3
discusses the proposed model in detail; Sect. 4 discusses the experimental results and com-
putational performance metrics, and Sect. 5 discusses the proposed work’s conclusion and
future directions.

2 Related work

In recent years, algorithms that use Deep Learning techniques to detect lung nodules
have been used a lot in medical research. Sunyi Zheng et al.[17] developed a deep learn-
ing model to locate lung nodules by taking into account the sagittal, coronal, and axial
slices of the CT images for the lung region is been evaluated. Here, the system is made
up of two parts. First, a supervised encoder-decoder is trained to find the nodule by
combining these slices. Multi-scale contextual information is extracted using 3D Dense
CNN to get rid of the nodules. Reza Majidpourkhoei et.al. [18], proposed A novel
deep-learning framework based on CNN to detect lung Nodules. This framework is
designed using light-footed CNN based on the LeNet-5 model. The lung nodule images
are processed and drawn on a patch basis. This model takes six hours to train, which
is a time-consuming process. Ying Su et.al., designed the framework for detection of
Lung Nodule detection using Faster R-CNN [19] and stated that optimizing the training
parameters like learning rate and batch size improves the accuracy of the detection and
also the dataset size is enhanced by including the medium lung nodule and by taking

13
 Multimedia Tools and Applications

into account the upper and lower nodules of the larger nodule that was identified on the
CT slice. In this design, the parameters have to be tuned manually.
Menglu Liu et.al. [20], proposed segmentation of the lung nodule using Mask-
R.CNN employs instance segmentation, this model is compared with the U-Net, and
Mask-R-CNN outperforms in segmentation. Linqin Cai et al., [21] demonstrated pulmo-
nary nodule detection based on the Mask-R-CNN and with a ray-casting volume render-
ing algorithm. where Mask-R-CNN helps to detect the pulmonary nodule by multiplying
the mask matrices and sequences of raw medical images, and the ray-casting technique
aids in visualizing the nodule in a 3D model. Here detecting the small nodule’s accuracy
has to be improved. These models are evaluated using LIDC-IDRI data sets, which is an
open-source dataset of Lung Images, and widely for research purposes.
The Deep Learning model uses the CNN architecture for automatic feature extraction
and diagnosis of Lung Images. CNN architecture consists of different hyper-parameters that
must be optimized to improve performance. In recent days different metaheuristic optimi-
zation techniques have been proposed. Wei-Chang Yeh et.al. worked on Simplified Swarm
Optimization (SSO) [22] combined with LeNet-5, where the author proposed the sequen-
tial Dynamic Variable Range (SDVR), in contrast to typical SSO the feasible range of the
next variable, which is determined by the present variable’s value. LeNet-SSO architecture
has improved the quality of the solution by tuning parameters and items. This system is
evaluated using MNIST, Fashion MNIST, and Cifar10 datasets. It outperforms when com-
pared with the other metaheuristic algorithms with LeNet. Singh et.al. proposed a Hybrid
MPSO algorithm, which uses multiple swarms in two levels to give a better solution for the
objective function [23]. the architecture of the CNN and hyper-parameters are optimized at
level 1 and level 2 respectively. To modify the exploration and exploitation characteristics
of particles and prevent the PSO algorithm from prematurely converging into a local opti-
mum solution, this technique employs sigmoid-like inertia weight. This system is evaluated
using different benchmark datasets like Cifar10, Cifar-100, MNIST, Covexset, and MDRBI
and it is outperformed when compared to randomly generated CNN.
Vijh et al., designed a hybrid bio-inspired algorithm [24] for automatic lung nodule
detection. Here a novel variant of whale optimization and adaptive PSO (WOA-APSO) is
used to optimize the feature selection and the selected features are grouped by employing
the linear discriminant analysis, which aids the reduction of dimensions spaces. the lung
images are enhanced using a wiener filter and by employing the different segmentation
techniques RoI of the Lung region is obtained, this system is evaluated using LIDC data-
sets, the accuracy, sensitivity, and specificity are 97.18, 97, and 98.66 respectively. The
PSO has the downside of easily falling into local optima in high-dimensional space and
having a slow convergence rate in iterative processes, despite being well suited for non-
linear complex problems. to avoid the pitfalls of premature optimization and to avoid being
stuck in neighborhood searches. we adopt another meta-heuristic approach called snake
swarm optimization for local search. Gunjan et.al., proposed work on analyzing different
metaheuristic algorithms namely Simulated annealing [25], Tree-of-Parzen estimator, and
Random search for optimizing the CNN structure hyper-parameter to classify the small
pulmonary lung nodules. Here the system uses the LIDC datasets, and the results show that
the SA performs well when compared to other metaheuristic algorithms. Sollini et al. [36]
explained the lung lesions classification using a deep learning algorithm. The two main
modules are the detection of lung nodules on CT scans and the classification of each nod-
ule into benign and malignant types. Computer Aided Diagnostics (CADe) and Computer
Aided Diagnostics (CADx) modules rely on deep learning techniques such as Retina U-Net
and Convolutional Neural Networks.

13
Multimedia Tools and Applications

3 Proposed lung nodule detection methodology

The main objective of the proposed methodology is to effectively detect the pulmo-
nary nodule from the CT lung images. To achieve this objective, in this paper we pro-
posed, an optimized MRCNN. The proposed MRCNN is enhanced by using a hybrid
particle snake swarm optimization algorithm ­(PS2OA). The hybrid ­PS2OA algorithm
is used to tune the hyper-parameter. The proposed approach consists of two main
stages namely, pre-processing and detection. The pre-processing is done by adaptive
median filter and CLAHE. After pre-processing the image is given to the Optimized
MRCNN, in which the nodule is detected. The structure of the proposed methodology
is given in Fig. 1.

3.1 Pre‑processing

The original Lung CT images are pre-processed to remove the noise and enhance image
contrast. For this, we apply Adaptive Right Median Filter (AMF) and Contrast Limited
Adaptive Histogram equalization (CLAHE) methods.

3.1.1 Adaptive median filter (AMF)

The purpose of using the Adaptive median filter is to reduce the distortion and preserve
the image edge details [28]. The advantage of utilizing an adaptive median filter over
a standard median filter is that the kernel size is adjustable in the area around the dis-
torted image, as a result, we can obtain better output and in contrast to the median filter,
it will not replace all of the pixel values with the median value. This algorithm works
on two levels.

Fig. 1  Workflow of proposed methodology

13
 Multimedia Tools and Applications

Level: 1 The first level involves determining the kernel’s median value.

P1 = ­Zmin- ­Zmed

P2 = ­Zmax- ­Zmed
If ­P1 > 0 and ­P2 < 0 go to Level 2
Else increase the kernel size
If kernel size <  = ­Smax iterate level 1
Else output Z ­ xy

Sxy—The local region of the gray level image at x,y.

Zmin, ­Zmax—Minimum and maximum gray level value in ­Sxy.
Zmed—Median gray level value in ­Sxy.
Zxy—Gray level coordinates at x,y
Smax—the maximum allowed size of the region ­Sxy.

Level: 2 In level 2 determine whether the current pixel value is an impulse (salt and pepper
noise) or not. If a pixel’s value is corrupted, it either modifies it using the median or keeps
the grayscale pixel value.

Q1 = ­Zmed—Zmin

Q2 = ­Zmed—Zmax
If ­Q1 > 0 and ­Q2 < 0 output ­Zxy
Else return ­Zmed

Here the original CT Lung image of size 224 × 224 is shown in Fig. 2 (a) and the S
­ max
maximum window size is assigned as 11. First, it is converted into a grayscale image as
depicted in Fig. 2 (b). Then the AMF method is employed on that image for denoising,
which is shown in Fig. 2 (c).

3.1.2 Contrast Limited Adaptive Histogram Equalization (CLAHE)

Histogram equalization is a method of processing images that modifies the intensity distri-
bution of the histogram to change the contrast of an image. CLAHE is the variant of Adap-
tive Histogram Equalization (AHE). It reduces amplified noise by limiting contrast ampli-
fication. It performs this by evenly spreading the portion of the histogram that exceeds the
clip limit across all histograms.
Histogram equalization (HE) is a technique that is often employed in image enhance-
ment approaches; however, HE raises contrast globally, [29] whereas AHE is a technique
that improves contrast in the local area. Unfortunately, AHE happens infrequently, which
raises the contrast. The CLAHE method can handle this by providing a clip limit that spec-
ifies the maximum height of a histogram and region size. Here the denoised images are
enhanced by having the clip limit as 0.02 and the tile size is assigned as 8X8. Clip limit
gives the contextual region of the CLAHE [30]. The Rayleigh distribution method is used
here to enhance the intensity values in every pixel. The bilinear interpolation method is

13
Multimedia Tools and Applications

Fig. 2  Pre-processing output

used to remove the artifacts near the boundary of the tiles. HE is based on a transformation
function, which is a combination of a probability distribution function (PDF) and a cumu-
lative distribution function (CDF). The general histogram stretching is given by Eq. 1,
( )
Omax − Omin
Pout = Pin − Imiin + Omin (1)
Imax − Imin
where ­Pout and P
­ in are the pixel value of the input image, I­min, ­Imax ­Omin, and O
­ max is the
input and output images’ respective minimum and maximum intensity levels.
( 2)
−x
PDFRayleigh = x − 2a2
2
e for x ≥ 0, a ≥ 0 (2)
a

where x is the intensity value of the input image and α Rayleigh distribution parameter.
Figure 3 depicts the Input CT lung image and Fig. 4 shows the image after applying the
CLAHE technique. Figures 5 and 6 show the plot of Histogram Equalization and CLAHE.

13
 Multimedia Tools and Applications

Fig. 3  Input CT Lung Image

3.2 Lung nodule detection and classification using optimized Mask RCNN

After pre-processing, the pre-processed images are sent into an optimized Mask RCNN
classifier to classify an image as malignant or benign. Mask R-CNN is a deep learning-
based approach that is mainly used for object detection and image segmentation. In many
computer vision tasks, including object detection, instance segmentation, and pose estima-
tion, the Mask R-CNN algorithm has been extensively used. The mask RCNN generates
the bounding box, segmentation mask, and corresponding class name. This works based
on the Feature Pyramid Network (FPN) and a ResNet101 backbone. To enhance the per-
formance of Mask RCNN, the hyper-parameter present in the Resnet is optimally selected.
For the parameter selection process, a hybrid optimization algorithm is presented. For
hybridization, particle swarm optimization and snake swarm optimization are presented.
The structure of Mask RCNN is presented in Fig. 7.
The Mask R-CNN model consists of three primary components which are the back-
bone, the Region Proposal Network (RPN), and RoIAlign. Backbone is a Feature Pyramid
network-style deep neural network that can extract multi-level image features. The ResNet
forms the backbone of the Mask R-CNN model. The CNN used here is ResNet 101. Fur-
ther, it has 3.8 × 109 floating point operations. The RPN uses a sliding window to scan the
input image and detects the infected regions in this study. The RoIAlign then examines the
RoIs obtained from the RPN and extends the feature maps from the backbone at various
locations. The RoIAlign is responsible for the formation of the precise segmentation masks
on the images. The RoIPooling in Faster R-CNN is replaced by a more precise and accurate
segmentation using the RoIAlign.

13
Multimedia Tools and Applications

Fig. 4  CLAHE Image

Fig. 5  HE plots of Input Lung CT

13
 Multimedia Tools and Applications

Fig. 6  Plot of CLAHE

3.2.1 ResNet 101 + FPN‑based feature map generation

ResNet-FPN is a backbone architecture used for feature extraction in Mask R-CNN. ResNet
is a deep convolutional neural network that is very effective for image classification tasks.
While FPN builds an in-network feature pyramid out of a single-scale input, it uses a top-
down architecture with lateral connections. In ResNet-FPN, FPN architecture is added on
top of the ResNet backbone to create a more effective feature extractor. The FPN compo-
nent allows for multi-scale feature maps to be generated from the input image, which can
improve object detection accuracy [14].
ResNet sets up a series of convolution, polling, and activation FC layers one after the
other. There are many types of ResNet architectures available, in this paper, ResNet 101 is
utilized. ResNet 101 consists of 101 layers. The proposed ResNet 101 has lower complex-
ity compared to VGG16 and VGG19 nets [11]. ResNet has three versions namely, ResNet
Version 1, ResNet Version 2, and ResNeXt. Each version has different characteristics.
As shown in Fig. 8, the ResNet-101 has a bottom-up path, which reduces the resolution
of the feature image. In contrast to ResNet-101, FPN improves the resolution of feature
images from the top down. Lateral links between ResNet-101 and FPN combine features
with the same resolution from ResNet-101 and FPN, respectively, to create new features in
FPN [10]. In this, two features with the same resolution from ResNet101 and FPN combine
to create a new feature in the path of FPN, and the ResNet101 backbone with FPN is used
to train models for Lung CT images. The ResNet-101 is trained to optimize the following
parameters;

• ResNet Version: ResNet 101 consists of a number of versions. The best version gives
the proper output. So, we select the optimal version for the segmentation process.
• Batch Size: We can select any size of batch for processing. This will affect the perfor-
mance. So, we choose the optimal batch size.
• Pooling type: Different types of pooling are available. So, we chose the optimal one.

13
Multimedia Tools and Applications

Fig. 7  Architecture of mask R-CNN

Fig. 8  ResNet-101 + FPN model

• Learning rate: For maximizing the final accuracy, the learning rate is another crucial
factor. It can be difficult to determine the proper learning rate.
• Optimizer: The optimizer is used in the fully connected layer.

The above-mentioned five parameters are optimally selected by using the ­PS2OA. The
range of overlay parameter configurations used for the P ­ S2OA algorithm is shown in Table 1.
Step 1: Solution encoding: Solution initiation is an important factor in the P ­ S2OA, which
is used to define the problem. In this paper, random initialization is used. Here, the param-
eters present in the ResNet101 namely, the version of ResNet, batch size, pooling type, learn-
ing rate, and optimizer are optimally selected by using the P ­ S2OA algorithm. Initially, these
2
parameters are randomly initialized. In the ­PS OA, the solutions are called swarms, and the
parameters are called particles. The initial solution format is given in Eq. (3).
{ }
Si = S1 , S2 , … ., Sn (3)

where Sn represent the nth swarm.

13
 Multimedia Tools and Applications

Table 1  Hyper-parameter range Hyper-parameter Range

Version [v1, v2, v3]

Batch size [32]
Pooling type [average, maximum, minimum]
Learning rate [0.1, 0.01, 0.001]
Optimizer [‘adam’, ‘rmsprop’, ‘sgd’]

Step 2: Fitness calculation: After the random creation, the fitness function is calcu-
lated for each swarm. In this paper, maximum accuracy is considered as the fitness func-
tion. The fitness function is given in Eq. (4).
Fitness = Max(Accuracy) (4)
Step 3: Update the solution using hybrid Particle Snake Swarm Optimization:
­ S2OA
After the fitness calculation, we update the solution. For updating, in this paper P
2
­ S OA is a combination of PSO and SSO.
algorithm is used. P

3.2.2 Particle Swarm Optimization

PSO is initialized based on fish schools and bird swarms in nature with a swarm intel-
ligence method. Every velocity vector and position vector in PSO is defined as a particle.
Each particle has its traverses and conducts a search space aligned with the best solution.
Particles already know the optimal location the complete particle swarm has computed.
The location and velocity vector updating process is formulated as follows,
( ) ( )
vik+1 = vki + c1 r1 pbestki − xik + c2 r2 gbest − xik (5)

xik+1 = xik + vik+1 (6)

3.2.3 Snake Swarm Optimization

Snake swarm optimization was developed in 2022 to reduce the mating characteristics of
snakes [35]. Mating is achieved when food is available and at low temperatures.
Stage 1: Initialization Phase:
The random initial population of the SSOA is presented as follows,
n
Nmale ≈
2 (7)

Nf = n − Nmale (8)
Here, Nf is defined as female individuals, n is defined as the number of individuals and
Nmale is defined as male individuals. The random initialization of the SSOA is divided into
two clusters such as male and female. In every iteration, the optimal individual candidate

13
Multimedia Tools and Applications

solution is computed by validating every group for optimal female and optimal male. The
food quantity and temperature are described as follows,
( −g )
t = exp (9)
t
(g − t)
fq = c1 exp (10)
t
Here, c1 is defined as constant to 0.5, t is defined as the total number of iterations, g is
defined as the current iteration. When fq < threshold , the snakes search for food by choos-
ing a random position and after that upgrade their position.
Stage 2: Male snake formulation:
To numerically design the exploration characteristics of the female and male snakes, it
is used,
xi,j (g + 1) = x(rand𝜖[1, n ],j) (g) ± c2 × ai,male (ub − lb) × rand𝜖U(0,1) + lb, Hereai,male
2
( )
frand,male (11)
= exp −
fi,male

Here, ± is a flag direction operator, fi,male is defined as the fitness of the male in the
group, frand,male is defined as the fitness of the earlier chosen random male snakes, ai,male is
defined as the capability to compute the food by male, rand is defined as the random num-
ber between 0 and 1, x(rand𝜖[1, n ],j) is defined as the position of a random male snake and xi,j
2
is defined as the male snake position.
Stage 3: Female snake formulation
xi,j = x(rand𝜖[1, n ],j) (g + 1) ± c2 × ai,female (ub − lb) × rand𝜖U(0,1) + lb, Hereai,female
2
( )
frand,female (12)
= exp −
fi,female

Here, ± is a flag direction operator, fi,male is defined as the fitness of the female in the
group, frand,female is defined as the fitness of the earlier chosen random female snakes,
ai,female is defined as the capability to compute the food by female, rand is defined as the
random number between 0 and 1, x(rand𝜖[1, n ],j) is defined as the position of a random
2
female snake and xi,j is defined as the male snake position. Female snake formulation is
given as
Stage 4: Exploration phase:
In the exploitation phase, two scenarios are considered for computing optimal solutions.
This condition is developed based on threshold parameters.

4 Condition 1:

FQ < threshold , it is updated by the below equation

( )
xi,j (g + 1) = xfood ± c3 × t × rand × xfood − xi,j (g) (13)

13
 Multimedia Tools and Applications

Here, c3 is equivalent to 2, xfood is defined as the position of the optimal individuals and
xi,j is defined as the position of individuals.

5 Condition 2:

FQ > threshold , it is updated based on the fighting and mating process.

Fighting Process.
The fighting capability of a female snake is formulated as follows,
( )
( ) −fbest,male
xi,j (g + 1) = xi,j (g) ± c3 × fi,female × rand × xbest,male − xi,f (g + 1) where fi,female = exp
fi
(14)
Here, fi,female is defined as the fighting capability of the female snake, xbest,male is defined
as the position of the best individual in the male group and xi,j is defined as the female
position. The fighting capability of the male snake is formulated as follows,
( )
( ) −fbest,f
xi,j (g + 1) = xi,j (g) ± c3 × fi,male × rand × xbest,female − xi,male (g) where fi,male = exp
fi
(15)
Here, fi,male is defined as the fighting capability of the male snake, xbest,female is defined as
the position of the best individual in the female group and xi,j is defined as the male position.

6 Mating process

In this phase, the female and male groups can upgrade their position,
( )
( ) −fi,male
xi,female (g + 1) = xi,f (g) ± MM i,female × rand × q × xi,male − xi,female (g + 1) , where MM i,female = exp
fi,female
(16)
( )
( ) −fi,female
xi,male (g + 1) = xi,m (g) ± MM i,male × rand × q × xi,female − xi,male (g + 1) , where MM i,male = exp
fi,male
(17)
Here, MM i,female is defined as the mating capability of a female, MM i,male is defined as
the mating capability of male, xi,m (g) is defined as the position of male and xi,f (g) is defined
as the position of female agents.

7 Proposed ­PS2OA

The major motive of the hybrid algorithm is to enhance the method’s ability to utilize PSO
while also exploring SSOA to achieve the optimization strength of both. The exploration
and exploitation of the SSOA were managed by the inertia constant in the hybrid algo-
rithm. Compared with the conventional computations, the primary agent’s location in the
hunting location is optimally upgraded. This is presented as follows,

13
Multimedia Tools and Applications

( )
( ) −fi,male
Female (mating) = xi,f (g) ± MM i,female × rand × q × xi,male − xi,female (g + 1) , where MM i,female = exp
fi,female
(18)
( )
( ) −fi,female
Male (mating) = xi,m (g) ± MM i,male × rand × q × xi,female − xi,male (g + 1) , where MM i,male = exp
fi,male
(19)
The location and velocity are adjusted to combined PSO and SSO variations and are
presented as follows,
( ) ( )
vik+1 = w.(vki +c1 r1 xi,female − xik ) + c2 r2 xi,male − xik (20)

xik+1 = xik + vik+1 (21)

Step 4: Termination criteria

The procedure is continued until the best hyper-parameter values are selected. The
selected value is given to the lung cancer detection process.

Algorithm 1  Pseudocode of the proposed hybrid algorithm

13
 Multimedia Tools and Applications

8 Region proposal network (RPN)

The RPN network utilizes the features extracted by ResNet101+FPN as input to generate
Regions of Interest (ROIs). In scenarios where the aspect ratios of objects differ, RPN can
predict both the foreground and background of an image. To efficiently generate candi-
date regions, the image box is positioned on the network, delineating the border box in the
expected feature image. A 3×3 convolutional layer scans the image, generating anchors
distributed across the image in different sizes. These anchors serve as starting points for
proposing potential regions of interest, facilitating subsequent object detection processes.
The network adapts its scale based on input images and utilizes a predefined set of anchor
boxes [15]. Each anchor corresponds to a unique bounding box and ground-truth class, allow-
ing for the recognition of defects of diverse sizes and shapes. Default bounding boxes encom-
pass a range of sizes and aspect ratios to accommodate various object characteristics. With
overlapping bounding boxes, determining the highest confidence score for detecting multiple
Regions of Interest (ROIs) becomes more straightforward [20]. This assessment is facilitated
by the Intersection over Union factor (IoU), calculated using equation (22), aiding in the
accurate identification of RoIs.
Area of Overlap
IoU = (22)
Area of union

9 ROI align model

RoIAlign processes a set of rectangular region proposals, extracting features from the feature
map corresponding to each proposal. In RCNN networks, pixel accuracy and the ability to dis-
tinguish individual branches within the same pixel target are crucial for mask branch detection.
After pooling and convolution of the original image, the image size undergoes changes, fol-
lowed by segmentation. Direct pixel-level segmentation techniques often fail to produce accurate
segmentation output. Hence, this paper proposes Mask RCNN, an enhanced version of Faster
RCNN. Additionally, CNN’s pooling layer is replaced with RoIAlign, which utilizes linear inter-
polation to preserve spatial details in the feature map. RoIAlign serves as a neural network layer
employed in object detection and instance segmentation algorithms, such as Mask R-CNN.
In Fig. 9, the green dotted lines are referred to as the 5X5 feature diagram, which is
derived after the convolution layer, and the feature corresponding to the ROI in the solid
line feature diagram is smaller, and RoIAlign maintains a floating-point number boundary
without scale processing. Initially, the feature’s small volume was separated into 2X2 units
(each unit boundary was not measured) and then each unit was separated into four smaller
units; the center point is illustrated as a four-coordinate position blue dot in the figure.
After that, two linear interpolations are performed to calculate the values of the four levels,
followed by maximum pooling or average voting to generate a 2×2 scale feature map.

10 Loss function

The loss function in Mask R-CNN, a popular instance segmentation model, is a compos-
ite function that combines classification, bounding box regression, and mask segmentation
losses. It serves to optimize the model parameters by minimizing the discrepancy between

13
Multimedia Tools and Applications

Fig. 9  Schematic diagram of the RoIAlign algorithm

predicted and ground-truth values for object classification, localization, and pixel-wise seg-
mentation simultaneously. This loss function plays a pivotal role in training Mask R-CNN
to accurately identify object instances and their corresponding masks in images. The loss
function of the proposed model is given in Eq. (23).
L(OMRCNN) = LossClass + LossBox + LossMask (23)
where; prediction loss of the presented class label is represented as Lossclass, the loss of
bounding box is represented as LossBox , and the presented segmentation mask loss is rep-
resented as Lossmask . The Lossclass is calculated based on the normal image and affected
image. The mathematical expression of LClass is given in equation (24).
( ) [ ( )( )]
Lossclass Ai , A∗i = −log Ai A∗i + 1 − Ai 1 − A∗i (24)

where, ­Ai represents the candidate anchor i target prediction probability of having a disease
and is the ground-truth label which is 1 for the positive anchor, otherwise 0. The below
equation describes the regression loss of the bounding box function;
( ) ∑ ( )
LossBox Bi , B∗i = SmoothL1 Bi − B∗i
(25)
i∈{x,y,w,h}

where;
{
0.5x2 , if |x| < 1
SmoothL1 (X) =
|x| − 0.5, otherwise (26)

predicted bounding box is defined as Bi, the GT based positive anchor is defined as B∗i .
The loss function of the mask is calculated as below;
[ ) ( )]
1 ∑ (
LossMask = − 2 Xij log Otij + 1 − Xij log 1 − Otij (27)
n 1≤i,j≤n

13
 Multimedia Tools and Applications

where Xij represent the value of a pixel (i, j) in a ground-truth mask of size n x
n and Otij is the predicted value of the same pixel in the mask learned for class
(t = 1 for Malignant and 0 for Benign).

11 Results and discussion

The experimental results obtained by the proposed approach are presented in this section.
The proposed method is executed in TensorFlow and performance is analyzed. The analy-
sis was executed on Google colab in “Keras 2.3.1” with the “TensorFlow 1. Upon which
system the experiment was conducted. “Windows 10” and had a “Random-Access Mem-
ory (RAM) of 8 GB” and “Graphics Processing Units (GPUs)” are used in this experiment.
The performance of the proposed approach is analyzed based on different metrics namely,
accuracy, precision, recall, and F-Measure.

11.1 Dataset description

Seven research institutions and eight private medical imaging businesses have collaborated
to create a public dataset called the Lung Image Database Consortium and Image Data-
base Resource Initiative (LIDC-IDRI) [26]. Ten hundred and eighty computed tomogra-
phy (CT) scans were included in the database, with slice thicknesses ranging from 0.6 mm
to 5.0 mm. Four radiologists read through these scans in two separate reading sessions.
First, radiologists identified potentially malignant lesions and divided them into three cat-
egories based on their size (nodules >  = 3 mm, nodules < 3 mm, and non-nodules). The
results from all four radiologists were compiled, and then each radiologist unblinded and
rechecked every annotation. In practical practice, detecting lung nodules requires scans
with thin slices. Therefore, scans with slice thickness greater than 2 mm were not consid-
ered. So the author included a total of 888 images included in for the analysis, after exclud-
ing those with inconsistent slice spacing [27]. According to NLST screening criteria, nod-
ules larger than 3 mm were judged to be significant lesions.

12 Experimental results

In this section, we presented the visual representation of the proposed experimental results.
Table 2 represent the visual representation of the detection output. In Fig. 10, we ana-
lyze the accuracy performance by varying epochs, and in Fig. 11, we analyze the loss s
by varying epochs. According to Fig. 15, we understand that as the number of epochs
increases, the loss value decreases.

12.1 Comparative analysis results

In this section, we compare our proposed work performance with different detection mod-
els namely, Faster RCNN (FRCNN), Single Shot MultiBox Detector (SSD), YoLo model
and SVM-based lung nodule detection.
In Fig. 12, the performance of the proposed approach is analyzed based on an accuracy
measure. When analyzing Fig. 12, the proposed method attained a maximum accuracy of

13
Multimedia Tools and Applications

Table 2  Visual representation of detection output, column (a) represent the input image, (b)represent the
Gray image, (c) represent the adaptive median filtered image and (d) detected output
Adaptive median
Input image Gray image Detected output
filtered image

Fig. 10  Epoch vs accuracy

13
 Multimedia Tools and Applications

Fig. 11  Epoch Vs loss

Fig. 12  Performance analysis based on accuracy

97.67% and ANN-based lung nodule detection attained an accuracy of 89%. Compared to five
existing classifiers, SVM-based classification attained the worst results. Due to hyper-parameter
optimization in MRCNN, our proposed method attained better results compared to the existing
techniques. In Fig. 13, the performance of the proposed approach is analyzed based on preci-
sion. A good classification should have the maximum precision value. When analyzing Fig. 13,
the proposed method attained the maximum precision of 95.7% which is 2.6% better than
FRCNN-based lung nodule classification, 4.5% better than SSD-based lung nodule classifica-
tion, 6.2% better than YoLo-based classification and 8.2% better than SVM based lung nodule
classification. The performance of the presented technique is analyzed based on recall is given
in given in Fig. 14. As per Fig. 14, we understand that ORCNN-based lung nodule classification
attained the maximum recall value compared to the existing techniques. Similarly, we attained
the maximum F-score value shown in Fig. 15. From the results, we can understand proposed
approach attained the maximum output compared to the existing techniques.

13
Multimedia Tools and Applications

Fig. 13  Performance analysis based on precision

Fig. 14  Performance analysis based on Recall

12.2 Comparative analysis with published work

To prove the efficiency of our proposed approach, we compare our work with already pub-
lished research works. For comparative analysis, we considered four research works namely
3DCNN [31], CNN [32], texture CNN [33], and BCNN [34]. These four techniques are deeply
explained lung nodule classification. So, we compare our research work with these papers.

13
 Multimedia Tools and Applications

Fig. 15  Performance analysis based on F-score

Table 3  Comparative analysis results

References Accuracy Recall precision F-score Dataset

MMEL-3DCNN [31] 90.6 83.7 - - LIDC-IDRI

CNN[ 32] 87.26 81 87.8 - LIDC- IDRI
Texture CNN [33] 90.91 91.39 90.46 94.14 MNIST dataset
BCNN [34] 91.46 91.94 - 93.35 LUNA16
Proposed 97.67 99 95.7 95.67 LIDC-IDRI

The comparative analysis result is presented in Table 3. In this paper, for lung nod-
ule classification, we optimized MRCNN. To improve the performance MRCNN classi-
­ S2OA algorithm. To prove
fier, the hyper-parameters are optimally selected using a hybrid P
the efficiency, we compare our work with [31–34]. When analyzing Table 3, our proposed
approach attained the maximum accuracy of 97.67% which is 90.6% for [31], 87.26% for
[32], 90.91% for [33], and 91.46% for [34]. Due to MRCNN and hyper-parameter optimi-
zation, our method produces superior classification outcomes.

13 Conclusion

In this section, the proposed methodology is effectively detecting the pulmonary nodule
from the CT lung images. To achieve this objective, in this paper we proposed, an opti-
mized MRCNN. The proposed MRCNN is enhanced by using a hybrid P ­ S2OA. The hybrid
2
­PS OA algorithm is used to tune the hyper-parameter. The proposed approach consists of
two main stages namely, pre-processing and detection. The pre-processing is done by adap-
tive median filter and CLAHE. After pre-processing the image is given to the Optimized
MRCNN, in which the nodules are detected. The performance of the proposed approach is
analyzed based on different metrics and effectiveness compared with state-of-the-art works.

13
Multimedia Tools and Applications

The performance analysis of accuracy is 94.67%, the recall value is 99%, the precision value
of the proposed method is 95.7% and the f-score value of the proposed method is 95.67%.
Overall, this integrated approach holds significant promise in enhancing the efficiency and
accuracy of lung cancer detection, thereby contributing to improved patient outcomes and
advancing the field of computer-aided diagnosis systems.

Authors’ contributions The manuscript was written through the contributions of all authors. All authors
have approved the final version of the manuscript.

Funding No Funds, grants, or other support was received.

Data availability Data will be made available on reasonable request.

Declarations
Ethical Approval Not Applicable.

Information consent Not Applicable.

Conflicts of interest The authors have no financial or proprietary interests in any material discussed in this
article. The authors declare that they have no conflict of interest.

References
1. Sung, H., Ferlay, J., Siegel, R. L., Laversanne, M., Soerjomataram, I., Jemal, A., & Bray, F. (2021). Global
cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185
countries. CA: a cancer journal for clinicians, 71(3), 209–249, doi: https://​doi.​org/​10.​3322/​caac.​21660.
2. Liu K (2022) Stbi-yolo: A real-time object detection method for lung nodule recognition. IEEE
Access 10:75385–75394. https://​doi.​org/​10.​1109/​ACCESS.​2022.​31920​34
3. Shariaty F, Mousavi M (2019) Application of CAD systems for the automatic detection of lung
nodules. Informatics in Medicine Unlocked 15:100173. https://​doi.​org/​10.​1016/j.​imu.​2019.​100173
4. He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask r-cnn. In Proceedings of the IEEE inter-
national conference on computer vision (pp. 2961–2969). https://​doi.​org/​10.​48550/​arXiv.​1703.​06870
5. Poojary, R., & Pai, A. (2019, November). Comparative study of model optimization techniques in
fine-tuned CNN models. In 2019 International Conference on Electrical and Computing Technolo-
gies and Applications (ICECTA) (pp. 1–4). IEEE. DOI: https://​doi.​org/​10.​1109/​ICECT​A48151.​2019.​
89596​81
6. Cui H, Bai J (2019) A new hyperparameters optimization method for convolutional neural net-
works. Pattern Recogn Lett 125:828–834. https://​doi.​org/​10.​1016/j.​patrec.​2019.​02.​009
7. Loussaief, S., & Abdelkrim, A. (2018). Convolutional neural network hyper-parameters optimiza-
tion based on genetic algorithms. International Journal of Advanced Computer Science and Appli-
cations, 9(10), doi: https://​doi.​org/​10.​14569/​IJACSA.​2018.​091031
8. https://​neptu​ne.​ai/​blog/​hyper-​param​eter-​tuning-​in-​python-​compl​ete-​guide
9. Gaspar, A., Oliva, D., Cuevas, E., Zaldívar, D., Pérez, M., & Pajares, G. (2021). Hyperparameter
optimization in a convolutional neural network using metaheuristic algorithms. In Metaheuristics in
Machine Learning: Theory and Applications (pp. 37–59). Cham: Springer International Publishing,
doi: https://​doi.​org/​10.​1007/​978-3-​030-​70542-8_2
10. Da Silva GLF, Valente TLA, Silva AC, De Paiva AC, Gattass M (2018) Convolutional neural network-
based PSO for lung nodule false positive reduction on CT images. Comput Methods Programs Biomed
162:109–118. https://​doi.​org/​10.​1016/j.​cmpb.​2018.​05.​006
11. Xue J, Shen B (2020) A novel swarm intelligence optimization approach: sparrow search algorithm.
Systems science & control engineering 8(1):22–34. https://​doi.​org/​10.​1080/​21642​583.​2019.​17088​30
12. Clerc, M. (2010). Particle swarm optimization (Vol. 93). John Wiley & Sons. https://​kamen​penkov.​
files.​wordp​ress.​com/​2016/​01/​pso-m-​clerc-​2006.​pdf
13. http://​www.​ijcse.​net/​docs/​IJCSE​13-​02-​05-​017.​pdf

13
 Multimedia Tools and Applications

14. Rere LR, Fanany MI, Arymurthy AM (2015) Simulated annealing algorithm for deep learning. Proce-
dia Computer Science 72:137–144. https://​doi.​org/​10.​1016/j.​procs.​2015.​12.​114
15. Kathpal S, Vohra R, Singh J, Sawhney RS (2012) Hybrid PSO–SA algorithm for achieving partitioning opti-
mization in various network applications. Procedia engineering 38:1728–1734. https://​doi.​org/​10.​1016/j.​pro-
eng.​2012.​06.​210
16. Chang, J., Li, Y., & Zheng, H. (2021). Research on key algorithms of the lung cad system based on
cascade feature and hybrid swarm intelligence optimization for mkl-svm. Computational Intelligence
and Neuroscience, 2021, DOI: https://​doi.​org/​10.​1155/​2021/​54910​17
17. Zheng S, Cornelissen LJ, Cui X, Jing X, Veldhuis RNJ, Oudkerk M, van Ooijen PMA (2021) Deep
convolutional neural networks for multiplanar lung nodule detection: Improvement in small nodule
identification. Med Phys 48(2):733–744. https://​doi.​org/​10.​1002/​mp.​14648
18. Majidpourkhoei R, Alilou M, Majidzadeh K, Babazadehsangar A (2021) A novel deep learning frame-
work for lung nodule detection in 3d CT images. Multimedia Tools and Applications 80(20):30539–
30555. https://​doi.​org/​10.​1007/​s11042-​021-​11066-w
19. Su, Y., Li, D., & Chen, X. (2021). Lung Nodule Detection based on Faster R-CNN Framework.
Computer Methods and Programs in Biomedicine, 200. https://​doi.​org/​10.​1016/j.​cmpb.​2020.​
105866
20. Liu, M., Dong, J., Dong, X., Yu, H., & Qi, L. (2018, September). Segmentation of lung nodule in CT
images based on mask R-CNN. In 2018 9th International Conference on Awareness Science and Tech-
nology (iCAST) (pp. 1–6). IEEE, DOI: https://​doi.​org/​10.​1109/​ICAwST.​2018.​85172​48
21. Cai L, Long T, Dai Y, Huang Y (2020) Mask R-CNN-Based Detection and Segmentation for Pulmonary Nod-
ule 3D Visualization Diagnosis. IEEE Access 8:44400–44409. https://​doi.​org/​10.​1109/​ACCESS.​2020.​29764​32
22. Yeh, W. C., Lin, Y. P., Liang, Y. C., & Lai, C. M. (2021). Convolution neural network hyperparameter
optimization using simplified swarm optimization. arXiv preprint arXiv:​2103.​03995, https://​doi.​org/​
10.​48550/​arXiv.​2103.​03995
23. Singh P, Chaudhury S, Panigrahi BK (2021) Hybrid MPSO-CNN: Multi-level particle swarm opti-
mized hyperparameters of convolutional neural network. Swarm Evol Comput 63:100863. https://​doi.​
org/​10.​1016/j.​swevo.​2021.​100863
24. Vijh, S., Gaurav, P., & Pandey, H. M. (2020). Hybrid bio-inspired algorithm and convolutional neural
network for automatic lung tumor detection. Neural Computing and Applications, 1–14, https://​doi.​
org/​10.​1007/​s00521-​020-​05362-z
25. Gunjan VK, Singh N, Shaik F, Roy S (2022) Detection of lung cancer in CT scans using grey wolf
optimization algorithm and recurrent neural network. Heal Technol 12(6):1197–1210. https://​doi.​org/​
10.​1007/​s12553-​022-​00700-8
26. Armato, S. G., McLennan, G., Bidaut, L., McNitt‐Gray, M. F., Meyer, C. R., Reeves, A. P., ... &
Clarke, L. P. (2011). The lung image database consortium (LIDC) and image database resource ini-
tiative (IDRI): a completed reference database of lung nodules on CT scans. Medical physics, 38(2),
915–931, doi: https://​doi.​org/​10.​1118/1.​35282​04
27. National Lung Screening Trial Research Team. (2011). Reduced lung-cancer mortality with low-dose
computed tomographic screening. New England Journal of Medicine, 365(5), 395–409, https://www.
nejm.org/doi/full/https://​doi.​org/​10.​1056/​nejmo​a1102​873
28. Verma K, Singh BK, Thoke AS (2015) An enhancement in adaptive median filter for edge preserva-
tion. Procedia Computer Science 48:29–36. https://​doi.​org/​10.​1016/j.​procs.​2015.​04.​106
29. Hana, F. M., & Maulida, I. D. (2021, February). Analysis of contrast limited adaptive histogram
equalization (CLAHE) parameters on finger knuckle print identification. In Journal of Physics: Con-
ference Series (Vol. 1764, No. 1, p. 012049). IOP Publishing, DOI https://​doi.​org/​10.​1088/​1742-​
6596/​1764/1/​012049
30. Ghani ASA, Isa NAM (2014) Underwater image quality enhancement through Rayleigh-stretching and
averaging image planes. International Journal of Naval Architecture and Ocean Engineering 6(4):840–
866. https://​doi.​org/​10.​1186/​2193-​1801-3-​757
31. Liu H, Cao H, Song E, Ma G, Xu X, Jin R, Liu C, Hung C-C (2020) Multi-model Ensemble Learning
Architecture Based on 3D CNN for Lung Nodule Malignancy Suspiciousness Classification. J Digit
Imaging 33:1242–1256. https://​doi.​org/​10.​1007/​s10278-​020-​00372-8
32. Agnes, S. A., & Anitha, J. (2020, January). Automatic 2D lung nodule patch classification using deep
neural networks. In 2020 Fourth International Conference on Inventive Systems and Control (ICISC)
(pp. 500–504). IEEE, doi:https://​doi.​org/​10.​1109/​ICISC​47916.​2020.​91711​83
33. Ali I, Muzammil M, Haq IU, Khaliq AA, Abdullah S (2020) Efficient lung nodule classification using
transferable texture convolutional neural network. Ieee Access 8:175859–175870. https://​doi.​org/​10.​
1109/​ACCESS.​2020.​30260​80

13
Multimedia Tools and Applications

34. Mastouri R, Khlifa N, Neji H, Hantous-Zannad S (2021) A bilinear convolutional neural network for
lung nodules classification on CT images. Int J Comput Assist Radiol Surg 16:91–101. https://​doi.​org/​
10.​1007/​s11548-​020-​02283-z
35. Hashim FA, Hussien AG (2022) Snake Optimizer: A novel meta-heuristic optimization algorithm.
Knowl-Based Syst 242:108320. https://​doi.​org/​10.​1016/j.​knosys.​2022.​108320
36. Sollini, M., Kirienko, M., Gozzi, N., Bruno, A., Torrisi, C., Balzarini, L., ... & Chiti, A. (2023). The
Development of an Intelligent Agent to Detect and Non-Invasively Characterize Lung Lesions on CT
Scans: Ready for the “Real World”?. Cancers, 15(2), 357, doi: https://​doi.​org/​10.​3390/​cance​rs150​20357

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

13