Multimedia Tools and Applications

https://doi.org/10.1007/s11042-022-13504-9

1218: ENGINEERING TOOLS AND APPLICATIONS IN MEDICAL

IMAGING

Convolution neural network based model to classify

colon cancerous tissue

Kusum Yadav 1 & Shamik Tiwari 2 & Anurag Jain 2 & Jalawi Alshudukhi 1

Received: 19 January 2021 / Revised: 5 September 2021 / Accepted: 13 July 2022

# The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract
Computer vision-based methods play a significant role in the recognition of cancerous
tissue from histopathological images. Therefore, computer-assisted diagnosis systems
provide an effective system for medical diagnosis. At the same time, conventional
medical image processing methods rely on feature extraction algorithms suited for a
particular problem. However, deep learning-based methods are becoming vital alterna-
tives with new developments in the machine learning area to reduce the complications of
the feature-based methods. Therefore, a Convolutional Neural Network-based model has
been proposed to categorize colon cancer histopathological images having multiple
classes of cancerous tissues. Eight different classes of cancer have been examined,
namely tumor epithelium, simple stroma, immune cells, complex stroma, normal mucosal
glands, debris, adipose tissue, and background. The Convolution Neural Network
(ConvNet) classification model is evaluated with four distinct activation layers, namely
Rectified Linear Unit (ReLU), Leaky Rectified Linear Unit (LReLU), Parametrized
Rectified Linear Unit (PReLU), and Exponential Linear Unit (ELU). Based on the
produced results, the ELU activation function has shown the highest classification
accuracy with on average 98% and in some cases 99%.

Keywords Cancerous tissue . Deep learning . Convolution neural network . Activation function

* Shamik Tiwari
[email protected]

1
Computer Science and Information Systems, College of Computer Science and Engineering,
University of Ha’il, Ha’il, Kingdom of Saudi Arabia
2
School of Computer Science, University of Petroleum and Energy Studies, Bidholi,
Dehradun 248007 Uttarakhand, India
 Multimedia Tools and Applications

1 Introduction

Computer vision-assisted approaches are widely applied to support medical specialists in

examining histopathological images for diagnoses and other purposes. Multiple cancers can
be detected using Computer Assisted Diagnosis (CAD) support systems, which rely heavily on
image processing [15, 16]. Cancer can grow in any part of the human body. It develops when
cells grow uncontrolled and dominate normal cells. It carries difficulties for the body to
function normally. Cancer has various types depending on the affected part of the body. It
can start in the colons, the breast, the lungs, or in the blood. Cancers are similar in some
manners, but they are dissimilar in the behaviors they mature and spread. A lump is referred to
as a tumor that is formed in all cases of cancers. However, all the tumors are not cancerous.
The medical specialist examines a part of the tumor to identify if there is cancer or not. Non-
cancerous tumors are termed benign, whereas cancerous tumors are termed malign. Few
cancers do not form any tumor, such as leukemia (blood cancer). It develops in the cells of
the blood. Cancer develops in the colon or the rectum, which is the last part of the human
digestive tract, referred to as colon cancer [18, 19]. Depending on where it grows, colon
cancers can also be termed rectal cancer or colorectal cancer. Due to the various common
features, colon cancer and rectal cancer are generally put together.
A better understanding of tumors can be provided by the quantitative and qualitative
analysis of various histopathology images of cancerous tissue. Moreover, they also explore
the multiple possibilities for cancer treatment. However, it is still a challenging job owing to
cellular heterogeneity [3, 21]. Although the manual diagnosis of histopathological slides by a
clinical practitioner is still essential, computer vision-assisted technology provides extraordi-
nary performance for examining tumor tissue [23].
The Sigmoid function was the most commonly used activation function, but when the
Rectifier Linear Unit (ReLU) was invented, it quickly became a better substitute for the
Sigmoid function due to its favorable influence on various machine learning tasks. It is
followed by introducing other ReLU activation functions, and this research investigated their
effects on the skin cancer dataset.
In this paper, the authors have utilized the ConvNet for classification. At the same, different
activation functions are examined for better results. Finally, the results of all of the activation
functions are evaluated and compared. This work could be of great help in the classification of
colon cancer cases.
This research paper is separated into six sections. Work done by other researchers is
explored in Section 2. Section 3 gives an overview of ConvNet and activation functions, the
proposed model, and the image dataset used for training and testing. Section 4 discusses the
simulation environment, performance measures, and results. In the end, Section 5 gives the
conclusion.

2 Literature review

In this section, the authors have reviewed the research work found in the domain of image-
based colon cancer classification. Work has been analyzed on the scale of methodology and
accuracy. Details are as follows:
Numerous physically interpretable and clinically important shape and morphology-based
features are fetched from segmented images which consist of gray levels texture features such
Multimedia Tools and Applications

as grey level cooccurrence matrix, color-based moments, features based on Law’s Texture
Energy, Tamura’s features, and wavelet features in the feature extraction phase. The wavelets
provide the multi-resolution-based study of microscopic images. Wavelet decomposition
offers compact features and provides better compression than other transforms such as Fourier
transform and Discrete Cosine Transform (DCT). The statistical features are calculated from
wavelet coefficients such as mean, standard deviation, kurtosis, entropy, contrast homogeneity,
energy, and sum for tissue categorization [12, 13, 20].
After successful feature extraction, various scholars have applied several supervised, semi-
supervised, and unsupervised machine-learning approaches to automatically detect tissue
categories in histopathological images. Support Vector Machine [11], K-Nearest Neighbour
[27], and Artificial Neural Networks [4] are classification models which are based on machine
learning and used for segmentation and categorization of histopathology images.
Nowadays, due to the accessibility of high computation power and enormous memory, it is
possible to utilize deep learning-based models like Convolutional Neural Network (ConvNet),
which provides leading-edge results in a large range of image categorization and object
recognition problems. In contrast to traditional machine learning models, which need custom-
ized feature extraction before training, ConvNet learns valuable features directly from the input
image by improving the learning function. Deep machine learning-based classification models
have recently produced outstanding results in a variety of image classification issues, including
multiple medical image investigations and, most notably, histopathological image analysis [5,
10].
ConvNet with the rectified linear unit as an Activation Function (AF) has been used in
numerous computer vision applications. ReLU has been utilized as a better match to biological
neural activation functions than the sigmoidal nonlinearity function. On the other hand, ReLU
has a weakness in that the gradient is zero when the unit is inactive or saturated. To eliminate
the potential problems due to the zero gradient, various other activation functions have been
suggested in the past [7].
Xu et al. [25] have utilized the deep convolutional activation features with Support Vector
Machine (SVM) for classification. The classification accuracy claimed by the authors was
97%, which is an acceptable result in the case of colon cancer. However, the work uses a long
process for classification. Besides, the used deep convolutional activation features are utilized,
as the name suggests, for more feature extraction than learning and decision making. At the
same time, there are no details about the used activation functions. Nevertheless, the authors
have shown the overall classification accuracy regardless of the accuracy of each Colon cancer
type.
Eckle, K. and Schmidt-Hieber, J [6] have shown the significance of activation function
while comparing deep network and linear spline methods. For this, the authors have utilized
the ReLU activation function. The authors have shown that deep neural networks with ReLU
activation function have performed better than spline methods through constructive proof.
Berner J. et al. [2] have demonstrated the pathologies that impede inverse stability and then
build a limited parametrization space. They achieved inverse stability for a Sobolev norm for
shallow networks. Authors have shown that by optimizing over such constrained sets, every
function that can be learned by optimizing over unconstrained settings can still be learned.
Though a lot of researchers have contributed to the domain of colon cancer classification,
the power of deep learning is still not utilized effectively to get more accuracy in image
classification. Therefore, authors have tried to use deep learning power in more accurate image
classification for colon cancer. Detail of the methodology is given in the next section.
 Multimedia Tools and Applications

3 Material and methodology

In this section, the authors have discussed the details of the convolution neural network,
activation function, proposed model, and image dataset utilized to train and test the model.

3.1 Convolution neural network (ConvNet/CNN)

Convolution Neural Network (ConvNet/CNN) is typically a specialized neural network that is

designed for computer vision applications. The structure of ConvNet is typically composed of
three different types of layers. Each of these layers has a separate rule for forwarding and
backward signal propagation [8].

3.1.1 Convolutional layer

A ConvNet generally comprises three layers:

& Convolutional layer.
& Sampling layer (pooling layer).
& Dense layer (fully connected layer).

These are shown in Fig. 2. The convolution layer performs convolution operation. The first
convolution layer receives an image as input. This operation is achieved on an input image
with a particular kernel (filter). A kernel is typically square, with a width ranging from three to
N pixels. A feature map is the result of the convolution layer. The feature map from the
previous convolutional layer is fed into the following convolutional layer. As a result, a feature
map is created by convolution the filter over the input’s individually identifiable elements.
The convolution process can yield a feature map of the different dimensions than input
subjective to the dimension of filter and padding preferences. It is essential to apply zero paddings
on input boundaries when the same dimension of output is required. In the case of zero paddings,
it has to add the required number of zero elements around the input boundaries. Sometimes, it is
required to reduce the dimension of input. The reduction of input can be amplified with the use of
stride. Stride gives an estimate of input points that are considered at the time of moving to
neighboring points during every step of convolution. If the dimension of stride is 2, the filter is
moved by 2 on every step and the resulting dimension of output is reduced by half.
Every convolutional layer consists of nonlinearity on output, which is also referred to as the
detector stage. Rectified Linear Unit (ReLU) is frequently utilized AF in ConvNet [22]. It is
identical to the activation function of the neural network.
The convolution layers are separated into two sublayers: The first sublayer (as shown in
Eq. (1)) is used to map the linear features which have been completed by performing fixed
dimension filters on the output of earlier layers.

bn ¼ wn  an1 ð1Þ
here ⊗ represents the mathematical convolution operator. The second sublayer (as shown in
Eq. (2)) of the convolutional layer is a non-linear mapping given as:
an ¼ f ðbn Þ ð2Þ
Multimedia Tools and Applications

3.1.2 Pooling layer

The pooling layer is not part of the learning process; however, it is used to decrease the
dimension of the feature map. The input is partitioned into several non-overlapping rectangular
windows and a single element out of several window elements selected as the output. As a
result, the feature map’s dimension is reduced yet the most important features are preserved.
The type of operation that is performed on each piece determines the nature of the pooling
layer. Examples of these operations are averaging, picking the maximal value, or perform the
linear combination. Max-Pooling is the most common form of pooling which operates by
choosing of maximal value [1]. The average pooling operation is given by Eq. (3):
1 Xn
xi ¼ y
i2wj j
ð3Þ
n
y is some feature in the window wj from features mapping.
n is the aggregate number of features in the window.

The max-pooling is given as in Eq. (4):

xi ¼ maxi2wj yj ð4Þ

3.1.3 Fully connected layers

The last layer of ConvNet is fully connected and is also referred to as Fully Connected Layer
(FCN). After the convolutional and the pooling operations, the feature map drives to this final
layer, i.e., fully connected layers. Generally, there is a minimum of one FCN layer in the
ConvNet. This layer performs the sum with weights to the previous layer’s output to decide a
particular target output result.

3.2 Activation function

Activation functions are an essential feature for all Artificial Neural Networks (ANNs). It
decides that a neuron should be triggered or not. Beyond this, whether the neuron receiving
information is appropriate for the given information or overlooked [26]. Two desirable
qualities of the perfect activation function are:

1. Creating a zero-centered distribution, which can aid in training efficiency

2. Saturation is one-sided, resulting in improved convergence

The first and second conditions are met by Leaky ReLUs and PReLU, while the second is not.
The second, but not the first, criterion is satisfied by vanilla ReLU. An Exponential Linear Unit
is an activation function that meets both conditions.

3.2.1 Sigmoid function

The sigmoid function is an activation function whose values are in the range [0, 1]. Its graphed
shape is similar to the Greek word sigma; that’s why it is called the sigmoid function. This
 Multimedia Tools and Applications

function has applications in neural networks. It is used in experiments where we have to

predict probability. The sigmoid function is given in Eq. (5)

1
f ð xÞ ¼ ð5Þ
1 þ ex
The derivative of the sigmoid is given in Eq. (6)

d
f ð xÞ ¼ f ð xÞð1  f ð xÞÞ ð6Þ
dx
Any real value can be used as the function’s input. The output ranges from 0 to 1. Between 0
and.25 is the derivative. The sigmoid function returns a value between 0 and 1. As a result, it’s
used as a binary class classification output layer. The vanishing gradient problem occurs when
the derivative value is less. As a result, sigmoid is rarely used for buried layers.

3.2.2 Rectified Linear Unit (ReLU) activation function

Another prominent activation function in deep learning models is the ReLU (Rectified Linear
Unit). When the input is negative, it returns zero, and when the input is nonnegative, it returns
the input as output [17]. This is given in Eq. (7).

0 if x 0
f ð xÞ ¼ ð7Þ
x if x  0

The function is depicted in Fig. 1(a). It has the gradient, which is shown in Eq. (8).

d 0 if x 0
f ð xÞ ¼ ð8Þ
dx 1 if x  0

The derivative in ReLU is not specified for input = 0, which is a problem. The key
disadvantage is that output 0 is produced for all negative input values. For positive input
values, it does not demonstrate the vanishing gradient problem. For negative input values, the
derivative vanishes.

Fig. 1 a ReLU, b Leaky ReLU/ Parametrized ReLu and c Exponential ReLu activation functions
Multimedia Tools and Applications

3.2.3 Leaky rectified Linear Unit (LReLU) activation function

This function is a variant of ReLU function, which returns the input as output when it is
nonnegative. Also, on negative input, it gives 0.01 times of the input as output [17], as shown
in Eq. (9). Leaky ReLU solves the problem of vanishing gradient for negative input values.

x if x > 0
f ð xÞ ¼ ð9Þ
0:01x if x  0

The function is depicted in Fig. 1(b). It has the gradient as shown in Eq. (10)

d 1 if x > 0
f ð xÞ ¼ ð10Þ
dx 0:01 if x  0

Negative inputs, as well as derivatives for negative inputs, are not discarded. It provides output
for negative inputs. The problem with leaky ReLU is that for input = 0, the derivative is not
defined.

3.2.4 Parametric rectified Linear Unit (PReLU) activation function

The parametric rectified linear unit function behaves similarly to the leaky ReLU, with the
only difference that the tunable leak parameter can be adjusted during the network learning.
The function can be defined as [9] in Eq. (11).

x if x > 0
f ð xÞ ¼ ð11Þ
x if x  0

The function is depicted in Fig. 1(b). Where α is a constant, it has the gradient as given in
Eq. (12).

d 1 if x > 0
f ð xÞ ¼ ð12Þ
dx  if x  0

3.2.5 Exponential Linear Unit (ELU) activation function

This activation function is also a variant of ReLU. It does not suffer from the dying ReLU
problem. The advantage of this activation function is that it gives a more accurate result and
the convergence rate is faster. Moreover, it is a unification between the good features of ReLU
and Leaky ReLU activation functions. The function is defined as [9] given in Eq. (13).

ðex  1Þ if x 0
f ð xÞ ¼ ð13Þ
x if x  0

Where α is a constant, the function is depicted in Fig. 1(b). It has the gradient as given in
Eq. (14)

d 1 if x > 0
f ð xÞ ¼ ð14Þ
dx f ð xÞ þ  if x  0
 Multimedia Tools and Applications

3.3 ConvNet Architecture

Three convolutional layers make up the ConvNet classification model developed in this study.
The filter size for the first convolution layer is 32  3  3, the second convolution layer is 64
3  3, and the final convolution layer is 128  3  3. Activation and maximum pooling layers
come after all of the convolution layers. Each of the three pooling layers has a 2  2 dimension.
The final layer is a dense layer that utilizes the softmax function as an activation function
with eight units since there is an 8 class labeling problem. The diagrammatic representation of
this model has shown in Fig. 2.
After the ConvNet classification model was created, an Adam optimizer is used during
compilation. Adam optimizer is the prominent regular optimization algorithm. Adam com-
bines the finest features of the AdaGrad and RMSProp methods to create an optimization
technique for noisy issues with sparse gradients. Adam is simple to set up, and the default
configuration parameters work well for most problems. The loss function utilized during the
learning process is categorical cross-entropy, which has the form given by Eq. 15

1 XN
T
eW yi xiþ byi
LCE ¼ log Pn W T x þb ð15Þ
N i¼1
j¼1 e
j i j

W = weight matrix; b = bias term; xi = ith training sample; yi = class label for ith training sample;
N = number of samples; W j and W yi are the jth and yi th column of W, respectively.
An additional layer is used to avoid overfitting during learning, referred to as the dropout
layer. The role of this layer is to randomly deactivate some neurons during the learning
process, which results in reducing the dependency on the learning set by a small degree. This
result can be observed by a gap between learning and validation accuracy and loss plots.

3.4 Dataset

Kather’s dataset [9] comprising of histopathological images of human colon cancer. It contains
seven types of sub-tissue organizations and 1 class for images with contextual backgrounds.
The pathology archive at the University Medical Center Mannheim provided ten anonymized

Fig. 2 An overview of ConvNet model for tissue classification problem

Multimedia Tools and Applications

H&E stained CRC tissue slides (Heidelberg University, Mannheim, Germany). This group
comprised both low-grade and high-grade tumors; no further selection was made. The slides
were digitized first. Following that, continuous tissue sections were manually labeled and
tessellated, resulting in 625 non-overlapping tissue tiles. The overall 5000 images are available
in this dataset with 625 images for every eight classes, namely tumor epithelium, simple
stroma, immune cells, complex stroma, normal mucosal glands, debris, adipose tissue, back-
ground. The original size of these images was 150 × 150 pixels; however, they are resized to
128  128 pixels. Few sample images from individual classes are shown in Fig. 3.

4 Simulation results

This section contains the detail of the simulation environment, various performance metrics, along
results found while classifying the images using the model proposed in the previous section.

4.1 Simulation environment

All experiments have been done using Keras deep learning library with Intel Core i3-
[email protected] GHz CPU on 64-bit Windows 10 OS.

Fig. 3 Sample images of image database separately for each class

 Multimedia Tools and Applications

4.2 Performance metrics

To evaluate the performance of CNN classification models, a confusion matrix is designed and
three statistical measures, precision, sensitivity, and F-score, are used. These metrics are
specified below [14, 24].

Precision (P) Precision represents the ratio of truly predicted positive observations of the total
positive predictions.

P ¼ Tþ =ðTþ þ Fþ Þ

Sensitivity (S) Sensitivity is the ratio of truly predicted positive samples to all samples in true class.

S ¼ Tþ =ðTþ þ F Þ

F-Score (F) F-Score or F1 is the weighted average of Precision and Sensitivity. F1 is typically
more valuable than accuracy when the class distribution is not even in a dataset. Hence, this
metric considers both false negatives and false positives for calculation.

F ¼ ð2 * P * SÞ=ðP þ SÞ

Accuracy (A) It is the fraction of the total perfection of the classification model and is
calculated as the sum of true labeling divided by the total observations.

A ¼ ðTþ þ T Þ=ðTþ þ T þ F þ Fþ Þ

Where:
& T þ and T  are the truly labeled positive and negative observations, respectively.
& F þ signifies the no of -ive samples labeled incorrectly as positive.
& F  signifies the no of + ive samples labeled incorrectly as negative.

4.3 Results

The performance of the proposed model under different configurations is shown in this
section. Details are as follows:

4.3.1 ReLU activation layers performance

In this simulation, original colored images are utilized for training and testing the ConvNet
classification model using ReLU activation layers as discussed in model architecture. The
Multimedia Tools and Applications

training of the ConvNet classification model continues up to 50 epochs. Dataset is divided into
two parts: learning and validation image sets. 70% of the image set is used for learning, and the
remaining 30% of images are utilized for validation. The complete image dataset is utilized as
a test set. The average precision, sensitivity, and f-score are 0.92, 0.92, and 0.92, respectively,
as shown in Table 1. The total accuracy attained from the trained ConvNet classification model
is 92.6. Figure 4 compromises the training & validation accuracy and training & validation
loss plots for the learning and validation samples. Learning curves are a common diagnostic
tool for deep learning systems that learn in steps. During training, we assess model perfor-
mance on both the training and hold-out validation datasets and plot the results for each
training step. During training, reviewing model learning curves can be utilized to diagnose
learning difficulties such an underfit or overfit model and an underfit or overfit model.

4.3.2 Leaky ReLU activation layers performance

In this simulation, original colored images are used for training and testing the ConvNet
classification model using Leaky ReLU activation layers, as discussed earlier. The training of
the ConvNet classification model continues up to 50 epochs. There are two sections to the
dataset: learning and validation image sets. 70% of the image set is used for learning, while
30% is used for validation. As a test set, the entire image dataset is used. Table 2 shows that the
average precision, sensitivity, and f-score are 0.93, 0.92, and 0.92, respectively. 92.3 is the
total accuracy of the trained ConvNet classification model. Figure 5 compromises the training
& validation accuracy and training & validation loss plots for the learning and validation
samples.

4.3.3 Parametrized ReLU activation layers performance

In this simulation, original colored images are utilized for training and testing the ConvNet
classification model using Parameterized ReLU activation layers, as discussed in earlier
sections. The ConvNet classification model is trained for up to 50 epochs. There are two
sections to the dataset: learning and validation image sets. 70% of the image set is used for
learning, while 30% is used for validation. As a test set, the entire image dataset is used.
Table 3 shows that the average precision, sensitivity, and f-score are 0.94, 0.94, and 0.95,
respectively. The total accuracy attained from the trained ConvNet classification model is 94.6.

Table 1 Performance of ConvNet classification model with ReLU activation layers

Texture classes Performance metrics

Precision Sensitivity f1-score

Tumour epithelium 0.91 0.88 0.90

Simple stroma 0.89 0.91 0.90
Complex stroma 0.90 0.88 0.89
Immune cell 0.88 0.98 0.93
Debris and mucus 0.96 0.82 0.88
Mucosal glands 0.94 0.99 0.96
Adipose tissue 0.98 0.92 0.95
Background 0.93 0.99 0.96
Average 0.93 0.92 0.92
 Multimedia Tools and Applications

Fig. 4 a Accuracy curves for simulation with ReLU activation layers (b) Loss curves for simulation with ReLU
activation layers

Table 2 Performance of ConvNet classification model with Leaky ReLU activation layer

Texture classes Performance metrics

Precision Sensitivity f1-score

Tumour epithelium 0.94 0.97 0.95

Simple stroma 0.88 0.93 0.90
Complex stroma 0.95 0.84 0.89
Immune cell 0.98 0.92 0.95
Debris and mucus 0.93 0.96 0.95
Mucosal glands 0.94 0.99 0.96
Adipose tissue 0.98 0.77 0.86
Background 0.81 1.00 0.90
Average 0.93 0.92 0.92

Fig. 5 a Accuracy curves for simulation with Leaky ReLU activation layers (b) Loss curves for simulation with
Leaky ReLU activation layers
Multimedia Tools and Applications

Table 3 Performance of ConvNet classification model with Parameterized ReLU activation

Texture classes Performance metrics

Precision Sensitivity f1-score

Tumour epithelium 0.92 0.96 0.94

Simple stroma 0.93 0.88 0.91
Complex stroma 0.93 0.87 0.90
Immune cell 0.93 0.97 0.95
Debris and mucus 0.92 0.96 0.94
Mucosal glands 0.98 0.98 0.98
Adipose tissue 0.98 0.95 0.96
Background 0.96 0.98 0.98
Average 0.94 0.94 0.95

Figure 6 compromises the training & validation accuracy and training & validation loss plots
for the learning and validation samples.

4.3.4 ELU Activation layers performance

In this simulation, original colored images are utilized for training and testing the ConvNet
classification model using exponential linear activation layers, as discussed in the earlier
section. The ConvNet classification model is trained for up to 50 epochs. There are two
sections to the dataset: learning and validation image sets. 70% of the image set is used for
learning, while 30% is used for validation. As a test set, the entire image dataset is used.
Table 4 shows that the average precision, sensitivity, and f-score are 0.96, 0.96, and 0.95,
respectively. The trained ConvNet classification model has a total accuracy of 95.8%. Figure 7
compromises the training & validation accuracy and training & validation loss plots for the
learning and validation samples.
After successfully implementing all four types of activation functions in the ConvNet
model, it is observed that the ELU activation function gives the highest accuracy. The accuracy

Fig. 6 a Accuracy curves for simulation with Parameterized ReLU activation layers (b) Loss curves for
simulation with Parameterized ReLU activation layers
 Multimedia Tools and Applications

Table 4 Performance of ConvNet classification model with ELU activation layers

Texture classes Performance metrics

Precision Sensitivity f1-score

Tumour epithelium 0.95 0.97 0.96

Simple stroma 0.96 0.91 0.94
Complex stroma 0.93 0.95 0.94
Immune cell 0.99 0.95 0.97
Debris and mucus 0.95 0.96 0.95
Mucosal glands 0.96 0.99 0.98
Adipose tissue 0.99 0.94 0.96
Background 0.94 1.00 0.97
Average 0.96 0.96 0.96

Fig. 7 a Accuracy plots for simulation with ELU activation layers (b) Loss curves for simulation with ELU
activation layers

achieved using ELU is 95.8%, which higher than others. Because learning can speed up by
centering the activations at zero, ELU, instead of batch normalization, employs the activation
function to attain mean zero. The bias shift in the ReLUs is what slows down the learning
process. Learning causes bias shifts in the following layers for those with mean activation
greater than zero. As a result, ELU is a viable alternative to ReLU. Although ELU takes longer
to compute, it compensates for this by achieving faster convergence during training. During
the test, however, ELU is slower to compute than ReLU.

5 Conclusions

This paper presented an analysis of cancerous tissue categorization using diverse activation
functions in the ConvNet classification model layers. This work recognizes the best appropri-
ate activation function for identifying the type of cancerous tissue. The performance of the
ReLU activation function and its variants, such as Leaky ReLU, ELU, and PReLU, has been
Multimedia Tools and Applications

investigated. The paper concludes that the exponential linear unit (ELU) as an activation
function is the most prominent for these types of images. It’s designed to combine the best
features of ReLU and the flaws of leaky ReLU — it doesn’t have the dying ReLU problem.
Still, it saturates for big negative values, effectively rendering them inactive. It can be
enhanced by establishing a further robust ConvNet classification model for better accuracy
and speed. This work can be expanded in the future by proposing a skin cancer detection
system based on mobile phones and IoT devices that uses machine learning and a handheld
camera or mobile camera to capture images of probable skin lesions and discern between
malignant and benign melanoma skin images. Moreover, the models can be Integrated with
domain-knowledge, such as combining the asymmetry, border irregularity, colour, diameter
and evolving size characteristics (well known as ABCDE” features) to design the more robust
recognition.

References

1. Agostinelli F, Hoffman M, Sadowski P, Baldi P (2014) Learning activation functions to improve deep
neural networks.arXiv preprint arXiv:1412.6830
2. Berner J, Elbrächter D, Grohs P (2019) How degenerate is the parametrization of neural networks with the
ReLU activation function?. arXiv preprint arXiv:1905.09803
3. Brouwer NP, Van der Kruijssen DE, Hugen N, de Hingh IH, Nagtegaal ID, Verhoeven RH, Koopman M,
de Wilt JH (2020) The impact of primary tumor location in synchronous metastatic colorectal cancer:
differences in metastatic sites and survival. Ann Surg Oncol 27(5):1580–1588
4. Damodharan S, Raghavan D (2015) Combining tissue segmentation and neural network for brain tumor
detection. Int Arab J Inf Technol (IAJIT) 12(1):42–52
5. Debelee TG, Schwenker F, Ibenthal A, Yohannes D (2020) Survey of deep learning in breast cancer image
analysis. Evol Syst 11(1):143–163
6. Eckle K, Schmidt-Hieber J (2019) A comparison of deep networks with ReLU activation function and linear
spline-type methods. Neural Netw 110:232–242
7. Jain A, Tiwari S (2021) Convolutional capsule network for COVID-19 detection using radiography images.
Int J Imaging Syst Technol 31:525–539
8. Jain A, Tiwari S, Sapra V (2019) Two-phase heart disease diagnosis system using deep learning. Int J
Control Autom 12(5):558–573
9. Kather JN, Weis CA, Bianconi F, Melchers SM, Schad LR, Gaiser T, Marx A, Zöllner FG (2016) Multi-
class texture analysis in colorectal cancer histology. Sci Rep 6(1):1–11
10. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, Van Der Laak JA, Van Ginneken B,
Sánchez CI (2017) A survey on deep learning in medical image analysis, vol 42. Medical image analysis, pp
60–88
11. Mukhopadhyay S, Kurmi I, Pratiher S, Mukherjee S, Barman R, Ghosh N, Panigrahi PK (2018) February.
Efficacy of hidden markov model over support vector machine on multiclass classification of healthy and
cancerous cervical tissues. In: Optical Diagnostics and Sensing XVIII: Toward Point-of-Care Diagnostics,
vol 10501. International Society for Optics and Photonics, p 105011M
12. Niu M, Yi M, Dong B, Luo S, Wu K (2020) Upregulation of STAT1-CCL5 axis is a biomarker of colon
cancer and promotes the proliferation of colon cancer cells. Ann Transl Med 8(15):1–20
13. Pavel MI, Sadique AM, Ritul RA, Khan S, Nath S (2017) Cancer detection using image processing
techniques based on cell counting, cell area measurement and clump detection (Doctoral dissertation,
BRAC Univeristy)
14. Powers DM (2020) Evaluation: from precision, recall and F-measure to ROC, informedness, markedness
and correlation. arXiv preprint arXiv:2010.16061
15. Rodríguez JH, Fraile FJC, Conde MJR, Llorente PLG (2016) Computer aided detection and diagnosis in
medical imaging: a review of clinical and educational applications. In: Proceedings of the Fourth
International Conference on Technological Ecosystems for Enhancing Multiculturality, pp 517–524
 Multimedia Tools and Applications

16. Setzer FC, Shi KJ, Zhang Z, Yan H, Yoon H, Mupparapu M, Li J (2020) Artificial intelligence for the
computer-aided detection of periapical lesions in cone-beam computed tomographic images. J Endod 46(7):
987–993
17. Shang W, Sohn K, Almeida D, Lee H (2016) Understanding and improving convolutional neural networks
via concatenated rectified linear units. In international conference on machine learning. PMLR, pp 2217–
2225
18. Siegel RL, Miller KD, Jemal A (2016) Cancer statistics, 2016. Cancer J Clin 66(1):7–30
19. Siegel RL, Miller KD, Fedewa SA, Ahnen DJ, Meester RG, Barzi A, Jemal A (2017) Colorectal cancer
statistics, 2017. Cancer J Clin 67(3):177–193
20. Su HH, Pan HW, Lu CP, Chuang JJ, Yang T (2020) Automatic detection method for cancer cell nucleus
image based on deep-learning analysis and color layer signature analysis algorithm. Sensors 20(16):4409
21. Thurmaier J, Heinemann V, Engel J, Schubert-Fritschle G, Wiedemann M, Nüssler NC, Ruppert R, Kleeff
J, Schepp W, Löhe F, Karthaus M (2021) Patients with colorectal cancer and brain metastasis: The
relevance of extracranial metastatic patterns predicting time intervals to first occurrence of intracranial
metastasis and survival. Int J Cancer 148(8):1919–1927
22. Tiwari S (2017) A pattern classification based approach for blur classification. Indonesian J Electr Eng Inf
(IJEEI) 5(2):162–173
23. Tiwari S (2018) An analysis in tissue classification for colorectal cancer histology using convolution neural
network and colour models. Int J Inform Syst Model Des (IJISMD) 9(4):1–19
24. Tiwari S, Jain A, Sharma AK, Almustafa KM (2021) Phonocardiogram signal based multi-class cardiac
diagnostic decision support system. IEEE Access 9:110710–110722
25. Xu Y, Jia Z, Wang LB, Ai Y, Zhang F, Lai M, Eric I, Chang C (2017) Large scale tissue histopathology
image classification, segmentation, and visualization via deep convolutional activation features. BMC
Bioinformatics 18(1):1–17
26. Xu B, Wang N, Chen T, Li M (2015) Empirical evaluation of rectified activations in convolutional network.
arXiv preprint arXiv:1505.00853
27. Ye J, Li T, Xiong T, Janardan R (2004) Using uncorrelated discriminant analysis for tissue classification
with gene expression data. IEEE/ACM Trans Comput Biol Bioinf 1(4):181–190

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