Noname manuscript No.

(will be inserted by the editor)

Colon cancer detection for Crohn’s disease based on transfer learning

approach with Caffe-MobileNetV2

B.Anil Kumar · Mohan Bansal

Received: date / Accepted: date

Abstract The leading cause of cancer-related death is colorectal (or) colon cancer. It begins as a normal
intestinal cell, progresses through stages of colon benign tissue and adenoma transforms into cancer, and
then becomes metastatic. A clinically relevant diagnostic need for early detection of colorectal cancer
(CRC) serves a purpose, to reduce the risk of overtreatment and associated complications. One of the
world’s leading causes of cancer death is human colon adenocarcinoma. Its development and outcome are
tightly regulated by a complex network of genetic connections. Histological image analysis is used to make
a definitive diagnosis of colon cancer. This research, utilized the deep learning (DL) technique to extract
features such as color, structure, and shape from a total of 10,000 colorectal cancer-related images to
achieve cancer classification and detection. With the advent of whole-slide imaging, deep neural network
(DNN) has been utilized for analyzing colon cancer with digital image processing (DIP). By combining
the DNN model with digital histopathological image pre-processing methods, our aim is to advance the
system for colon adenocarcinoma and benign tissue classification and detection. The goal is to classify and
detect images and extract features using the Caffe-MobileNetV2 (CMNV2) design. Colon cancer detection
has been performed using Caffe model and classification is performed using the modified MobileNetV2
model. In this research, added our five more layers to the pre-trained MobileNetV2 architecture to reduce
training data for colon cancer diagnosis and improve classification accuracy. Experimental results using
histopathological images showed that proposed approach works successfully with accuracy of 99.95%.
Other performance metrics such as 99.90% precision, 100% recall, 99.95% f1-score, and 0.05% error rate
of our CMNV2 model indicate that it is more effective and outperforms the previous models.
Keywords Colon Adenocarcinoma, Colon Benign Tissue, Deep Neural Network · Caffe Model ·
Modified MobileNetV2 Model · Transfer Learning · Colon Cancer Classification · Colon Cancer
Detection.

1 Introduction

Colorectal cancer is a third most frequently diagnosed cancer and responsible for the majority of cancer-
related deaths Siegel et al. (2019). By 2030, 2.2 million new cases of colorectal cancer infections and 1.1
million deaths are expected worldwide Arnold et al. (2017). Colon adenocarcinoma (COAD), it’s one
common type of colorectal cancer, which primarily affects lining of the colon. The majority of the time,

B.Anil Kumar
Research Scholar
School of Electronics Engineering
VIT-AP University, Amaravati, India
E-mail: anilkumar.20phd7071@vitap.ac.in
Mohan Bansal
Sr. Assistant Professor
School of Electronics Engineering
VIT-AP University, Amaravati, India
E-mail: mohan.bansal@vitap.ac.in
2 B.Anil Kumar, Mohan Bansal

COAD transfers to nearby organs after entering the intestinal lumen. Being a highly aggressive malig-
nancy, it carries a very high risk of mortality and complications. Despite advances in surgical methods,
radiation, and chemotherapy as clinical treatments for colon cancer, patients still inadequate diagnosis
and treatment. Numerous regulatory mechanisms such as deoxyribonucleic acid (DNA), methylation, hi-
stone deacetylation, and micro ribonucleic acids (miRNAs) expression influenced genes to increase cancer
incidence and growth Perri et al. (2017). Furthermore, in order to increase our understanding of can-
cer and disease severity, it is crucial to uncover novel regulatory factors and targets for COAD therapy.
Maintaining cellular metabolism, including regulation of maturation, transport, stability, and destruction
of various forms of RNA, relies on post-transcriptional gene regulation (PTGR). RNA-binding proteins
(RBPs), is essential for the mechanisms involved in PTGR. They interact with the mRNAs of target
genes and control how those mRNAs are processed to affect cell behavior. Indeed, formation of ribonu-
cleoprotein (RNP) complexes with various nuclear RBPs regulates each of these actions Masuda and
Kuwano (2019). Recent studies have associated RBP with various malignancies, developmental disor-
ders, diabetes, muscle wasting, and neurological disorders De Conti et al. (2017). Colon cancer is greatly
affected by RBP. The RBP member RNA-binding motif3 (RBM3) is upregulated in a step-dependent
way, so its overexpression may result in the development of cancer. Increased CUGBP elav-like family
member1 (CELF1) expression drives intestinal epithelial cells into the growth-arrested G1 phase, whereas
the absence of CELF1 promotes cell division. An aggressive phenotype, decreased survival and increased
tumor recurrence are associated with lin-28B (LIN28B) overexpression. These studies are insufficient to
fully explore the spectrum of RBPs in COAD. Crohn’s disease (CD) and ulcerative colitis are two of the
more common clinical manifestations of inflammatory bowel disease (IBD), which is caused by chronic
gastrointestinal inflammation that is not infectious. IBD is much more common today than it was decades
ago. There were 6.8 million cases worldwide Alatab et al. (2020). Studies have associated IBD with a
higher risk of colorectal cancer Jung et al. (2017). This is most likely due to persistent and continuous
intestinal inflammation. In general, identification of biomarkers in liquid biopsies such as Identification
and characterization of circulating tumor cells (CTCs). This provides an effective and practical method
for early cancer detection and the detection of the presence of minimal residual disease Hendricks et al.
(2021). The Purpose of CTC detection and clinical implications for patients having cancer diagnosis has
long been established Pantel and Alix-Panabières (2019). Various technical systems for CTC analysis
have been developed over the past decades, each with different levels of sensitivity and specificity. In
general, following he can distinguish two core ideas. Primarily used polymerase chain reaction (PCR) to
indirectly detect her CTCs, a molecular induction strategy Hinz et al. (2017), Hendricks et al. (2018). An
immunostaining cytological approach to directly identify and quantify CTCs Cieślikowski et al. (2020),
Abdalla et al. (2021). Using a sensitive and specific real-time PCRquantitative PCR (RTPCR-qPCR)
technique to measure the expression of cytokeratin20 (CK20) in mononuclear cells (MNC). CK20 is a
patient-specific independent indicator of poor diagnosis with CRC is already possible. Furthermore, stud-
ies have indicated that CK20 is a predictor of response to the neoadjuvant chemoradiotherapy Patients
having rectal cancer, but not a prognostic factor Hendricks et al. (2021). Furthermore, the observation
is suchlike the indicated molecular method of CK20 expression has been detected in blood samples from
CRC patients may help in the early diagnosis of recurrence Hinz et al. (2017), Hendricks et al. (2018).
As a result of this finding, can hypothesize that CK20 may also be effective in the early detection of
CRC. Almost all colon and rectal malignancies are caused by adenocarcinoma which causes the rectum
and large intestine colon‘s linings damage by this condition. Polyp-like growths, known as colon cancer,
usually form on the sur of the lining of the colon or rectum Makhdoomi et al. (2020). The rectum or
colon can invade nearby or nearby lymph nodes. Family medical history and several dietary factors, such
as insufficient fiber and too much fat all increase the risk of colorectal cancer. Bloody stools, fatigue, and
weakness are typical symptoms. As a result, early cancer becomes more treatable. A number of computer-
aided diagnostic systems (CAD) have recently been developed to automatically monitor tumor or cancer
growth in the colon Kumar et al. (2022), Yildirim and Cinar (2022). The identification of colon cancer
using techniques of deep learning and machine learning method from analyzing histopathological images
are two examples of approaches related to cancer detection of artificial intelligence (AI). Recent research
in this area using machine learning has shown positive results in preventing colorectal cancer. Machine
learning-based cancer diagnostic methods leveraging big data are more complex and time-consuming as
they require manual feature detection and proprietary classifiers for detection Hammad et al. (2021).
However, most of the problems of previous machine learning techniques have been solved by deep learn-
ing methods Hammad et al. (2020). These methods use deep models to integrate the process of feature
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 3

extraction and classification into a single step. Researchers used transfer learning techniques, pre-trained
models, or their own deep models when using system-based deep learning approaches. Convolutional neu-
ral networks (CNNs), which have achieved high accuracy in colon cancer detection, are the most popular
deep learning models for analyzing 3D ultrasound imaging, detection of COVID and other diagnostics
models for detecting cancer Hammad et al. (2022).

The work’s major contribution is:

i. The histopathological image dataset is used for training and testing our CMNV2 design approach for
colon cancer detection.
ii. Five layers are added to the MobileNetV2 model to create a specific model architecture using a
transfer learning approach.
iii. Caffe model and modified MobileNetV2 output classification utilized to colorectal cancer detection.
ii. For improving CMNV2 model accuracy, developed a modified MobileNetV2 design. Added five own
layers and created our own architecture for the proposed model, as shown in 1 by utilizing minimum
model parameters.
The rest of the paper is categorized as follows. Section 2 describes related work. The implications of
conventional MobileNetV2 and modified MobileNetV2 for different convolutional methods are discussed
in Section 3. Section 4 details our proposed method for detecting colon cancer using the improved
MobileNetV2 and Caffe models. Section 5 describes how the proposed methodology works when used
with confusion matrices and various parameters. Section 6 is discussed about results, including proposed
model accuracy as well as the detection of colon adenocarcinoma and benign tissue. Section 7 concludes
and summarizes the paper’s theory.

2 Related work

Colon and rectal cancer, the second most common cancer in wealthy civilizations, is rare in developing
countries. Colorectal cancer affects more than 940,000 people worldwide each year and kills about 500,000
Who (2020). Due to population growth, the prevalence of malignant tumors is increasing globally. Al-
though this could affect people of all ages, it is most common in people aged 50 to 60 Arslan et al. (2018).
Digital image processing (DIP) with deep learning (DL) a new approach was developed to detect five
different kinds of colon as well as lung tissue, including 2 benign and 3 malignant. These are identified
with a classification framework by examining histological images Masud et al. (2021). A convolutional
neural network (CNN)-based deep learning methodology was developed for effective colon cancer diag-
nosis on histopathological images Sakr et al. (2022). Deep convolutional neural network (DCNN) design
using pre-processing technique is constructed together with digital image processing (DIP) as well as
modern-deep learning (MDL) for identification of colon cancers on digital histopathology images Hasan
et al. (2022). Computer vision-related deep learning techniques for identifying colorectal cancer have
been the basis of a survey on polyp recognition, localization, and segmentation methods Sanchez-Peralta
et al. (2020). Dual-tree and double-density 2-D wavelet transform using a random forest classifier model
were developed to detect colon cancer in colon biopsy images, obtained from Aster medcity at various
magnifications Babu et al. (2018). A convolutional neural network is trained with binarized weights
using colonoscopy frames and frame labels as input data for a new classification. Using the Asu Mayo
test clinic database containing colonoscopies of various patients. Reduced CNN and made it suitable for
use on medical hardware using binarized weights and kernels Akbari et al. (2018). A Fourier transform
infrared (FTIR) technique is presented to overcome the similarity between cancer and healthy patients.
This is done by using the area and height ratios of signals and a number of statistical features. This
study demonstrated how to combine multiple features of plasma FTIR spectra to distinguish between
colon cancer patients and healthy individuals Toraman et al. (2019). An image data set (LC25000) was
created to contain images of histological entities such as cancer pathology. To address this properly train
a computer model, machine learning techniques require large data sets Borkowski et al. (2019). Deep
learning approach was developed for histopathological images detection of colon cancer. This work has
three CNN variants of ResNet-18, ResNet-34 and ResNet-50 are used for analyzing digital images of
pathological and evaluating their applicability Bukhari et al. (2020). An Convolutional neural networks
is presented as computer-aided diagnostic technique. This model is used to detect malignancy of the squa-
mous cell, lung and colon adenocarcinoma for analyzing digital pathological images of such cancer types
4 B.Anil Kumar, Mohan Bansal

Mangal et al. (2020). Classification of lung and colon cancer tissue images was performed using artificial
intelligence-supported models and optimization techniques. The effective features captured through the
two optimization techniques used are merged, categorized using support vector machine (SVM) tech-
nique, and trained using the DeepNet-19 model Toğaçar (2021). convolutional neural networks (CNNs)
like State-of-the-art such as AlexNet, visual geometry group (VGG), ResNet, DenseNet, and Inception
models is designed are reviewed and compared. Colon cancer regions are classified and highlighted us-
ing a deep learning architecture in the context of minimally annotated histopathological data Hamida
et al. (2021). As a method, the automatic identification of eight different tissue types recognized in the
histological evaluation of CRC has been proposed. The CNN topologies is utilized for transfer learn-
ing. The modified structural component of CNNs extracts features from images and feeds them into
popular machine learning algorithms such as naive Bayes, multi-layer perceptrons, k-nearest neighbors,
random forests, and support vector machines Ohata et al. (2021). A faster regional convolutional neural
network (RCNN) model had developed, which automatically identify budding regions based on patho-
logical sections to aid in the clinical diagnosis and treatment of colorectal cancer Lu et al. (2022). Using
age period cohort (APC) modeling, world health organization (WHO) mortality database of mortality
data from 1989 to 2016. It is used to estimate colon mortality in 42 countries by 2035, which predicts
death from cancer and rectal cancer Araghi et al. (2019). A DCNN tecnique was created for identifying
cancer-related tissue components in colon cancer. Later tested this approach using a cancer genome at-
las (TCGA) as well as pathological images belong to patients with colorectal cancer. The peritumoral
stroma (PTS) score of LNM was evaluated for its predictive power Kwak et al. (2021). A deep learning
strategy is designed to distinguish between normal mucosa, early preneoplastic lesions, adenoma, and
cancer at four stages in malignant tissue development Sena et al. (2019). A CNN model was developed
to predict colon cells imaging data, creating a new diagnostic tool. Colon cells image classification is
performed using a CNN with maximum and average pooling layers and a MobileNetV2 model Tasnim
et al. (2021). Pre-trained convolutional neural networks are examined using a variety of commonly avail-
able like squamous cell carcinoma, benign tissue as well as adenocarcinoma lungs. Data are evaluated
using six state-of-the-art off-the-shelf CNN designs, including VGG-19, Alex Net, ResNet-18, ResNet-
34, ResNet-50, and ResNet-101 Abbas et al. (2020). Pre-trained CNN models based on eight different
techniques: VGG16, NASNetMobile, InceptionV3, InceptionResNetV2, ResNet50, Xception, MobileNet,
and DenseNet169 are proposed to identify lung and colon cancers using histopathological images Garg
and Garg (2020). Different transfer learning methods have been used in studies comparing over typical
transfer learning model’s performances. These three separate datasets, Amazon Reviews, Reuters-21578,
and Office-31, are used to run the model. It emphasizes the importance of choosing the best transfer
learning model for different real-world applications Zhuang et al. (2020).

3 MobileNetV1 vs MobileNetV2 vs Modified MobileNetV2 vs proposed CMNV2

A modified MobileNetV2 method is utilized as such a transfer learning approach, with a Caffe approach
served like a single-shot detector (SSD) during detecting as well as verification. A better variant of
MobileNetV1 architecture classifier is MobileNetV2 architecture classifier. Compared to MobileNetV1,
the MobileNetV2 model uses an inverse residual block with a linear bottleneck. Compared to the original
MobileNetV1, it has significantly lesser parameters. Higher image dimensions give good results, where
MobileNetV2 accepts sizes larger than 32 × 32 as a input.
MobileNetV1 is a convolutional neural network architecture. The main goal of MobileNetV1 is to
provide a lightweight and efficient neural network architecture that can run on various devices with
limited computing resources. This is achieved by using a combination of depthwise separable convolutions
and pointwise convolutions. Depthwise separable convolutions are a type of convolutional operation that
factorizes a standard convolution into two separate operations: a depthwise convolution and a pointwise
convolution. This reduces the computational cost of the convolutional operation while maintaining the
model’s accuracy. Pointwise convolutions are standard convolutions with a 1 × 1 filter, which are used
to change the number of feature channels in a feature map. MobileNetV1 is composed of a series of
depthwise separable convolutions and pointwise convolutions, followed by a global average pooling layer
and a fully connected layer. The architecture also includes a technique called “depth multiplier”, which
controls the number of channels in each layer of the network, allowing the user to trade off model size
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 5

and accuracy. Overall, MobileNetV1 is an efficient and effective neural network architecture for various
vision applications, and has been widely used in various real-world applications.
MobileNetV2 is a follow-up architecture to MobileNetV1, designed to improve the performance and
efficiency of neural networks. The main innovation in MobileNetV2 is the use of inverted residual blocks,
which consist of a lightweight bottleneck block followed by a linear layer, and a shortcut connection that
bypasses the block. The bottleneck block applies a 1 × 1 pointwise convolution to reduce the number of
input channels, followed by a depthwise separable convolution, and another 1 × 1 pointwise convolution
to increase the number of output channels. The linear layer then applies a pointwise convolution to adjust
the number of channels to the desired value. The inverted residual block is called “inverted” because it
expands the number of channels first and then reduces them, which is opposite to the traditional residual
block used in other architectures. This approach allows MobileNetV2 to use more non-linearity at lower
computational cost, leading to better performance with similar model size. MobileNetV2 also includes
several other optimizations such as linear bottlenecks, which reduce the dimensionality of feature maps
and increase the efficiency of convolutional layers. The architecture also uses a feature map reduction
block to reduce the resolution of feature maps, and a new form of batch normalization that improves the
training stability. Overall, MobileNetV2 improves on the efficiency and performance of MobileNetV1,
and has become a popular choice for mobile vision applications that require real-time performance and
low power consumption.
Caffe is a deep learning framework that allows researchers and developers to create and deploy convo-
lutional neural networks. MobileNetV2 is a convolutional neural network architecture designed for various
devices that provides a good trade-off between model size, speed, and accuracy. To use MobileNetV2
in Caffe, needed to define the network architecture in a Caffe-compatible format, and then train the
network on a suitable dataset. Caffe provides a set of tools for data preparation, network definition,
and training, which can help to create a MobileNetV2 model that meets our specific requirements. This
network definition creates an input layer with a shape of (1, 3, 224, 224), which is the input format used
by MobileNetV2. It then defines a convolutional layer with 32 output channels, a kernel size of 3 × 3,
and a stride of 2, which is followed by a series of inverted residual blocks. To train this network on a
suitable dataset, you would use the Caffe tools to prepare the data, set the training parameters, and
start the training process. After the training is complete, you can use the resulting MobileNetV2 model
for inference on new data. Overall, MobileNetV2 can be used in Caffe to create efficient and accurate
convolutional neural networks for different vision applications.
Fig. 1 shows a architecture of the modified MobileNetV2 architecture. Similar to the convolutional
neural network-based deep learning approach, the technique makes use of layer upon layer including con-
volution, pooling, dropout, nonlinear, fully connected as well as linear bottlenecks. Design has seventeen
3 × 3 convolutional, max pooling average, a classification as well as a 1 × 1 convolutional layers. To train
and classify a colon cancer detection system, a sufficiently large dataset is required.
The proposed approach aims to increase the accuracy of colon cancer detection, by utilizing lesser
parameters for trainable. A key building block is the depth-wise separable convolutional layer (DWSC)
that makes MobileNetV2 extremely fast. To evaluate how well the proposed technique performs on various
image dimensions, our modified MobileNetV2 model is constructed for its 224 × 224, 192 × 192, 160 × 160,
and 128 × 128 images. These were used to compare the modified MobileNetV2 image dimensions to the
previous MobileNetV1 method image dimensions including the Caffe model to select the best model
Kumar and Bansal (2023).

Table 1: Architecture of the Proposed modified MobileNetV2

Layer Name Output Size Count of Parameters used

AveragePooling2D (1,1,1280) 0
Flatten (1280) 0
Dense (256) 327936
Dropout (256) 0
Dense 1 (2) 514
Total Parameter: 2,586,434
Trainable Parameter: 3,28,450
Non-trainable Parameter: 2,257,984
6 B.Anil Kumar, Mohan Bansal

Fig. 1: Modified MobileNetV2 Model

Added our own 5 layers such as 7 × 7 size average pooling, flattening, dense with 256 neural networks,
0.5 dropout size and dense 1 of 2 for detection into the architecture which changed and improved the
standard model of MobileNetV2. The remaining layers also are utilized for feature extraction in addition
with these layers. Out of the total 25,86,434 parameters of the proposed CMNV2 model, the improved
MobileNetV2 model Kumar and Bansal (2023) only requires 3,28,450 trainable and 22,57,984 untrainable
parameters. With comparison, the standard or regular MobileNetV2 model requires 34,112 untrainable
parameters and 22,23,872 parameters as trainable with a total of 22,57,984 parameters. As a result,
improved version like MobileNetV2 architecture has fewer trainable parameters and produces results
much faster than his original MobileNetV2 model. The dimensions, shapes, amount of parameters for
every layer as well as building block are displayed in Table 1.

Table 2: Different MobileNet models Comparison with CMNV2 model

Sl.No. Model Epoch Batch Size Total Run Time Parameters used Accuracy (%)
1 Standard MobileNetV1 60 48 40,464 Seconds 42,33,978
2 Standard MobileNetV2 60 48 32,400 Seconds 35,04,872 89.35
3 MobileNetV2 without Caffe 60 48 10,800 Seconds 3,28,450
4 CMNV2 (Proposed Model) 60 48 10,800 Seconds 3,28,450 99.95

Performance of the modified MobileNetV2 without Caffe, the standard MobileNetV1, the standard
MobileNetV2, and our suggested CMNV2 architecture are all displayed in Table 2.

4 Proposed Method

The CMNV2 model, which is utilized for colon cancer diagnosis using histopathology images, is the
subject of the suggested methodology. In the proposed approach, the design of our model is analyzed
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 7

and each layer’s weights are stored in two sets of files. The layer model architecture is contained in the
prototxt file as the first, and the weights for each layer are kept in the Caffe model as the second. The
CMNV2 model was developed with DNN architecture utilizing TensorFlow, Keras, and OpenCV libraries,
and it utilizes a method to detect colon adenocarcinoma and benign tissue. After each convolutional layer,
the neural network provides nonlinearity using a activation function like rectified linear unit (ReLU).
For all positive outcome values from layers of convolutional, ReLU like function is linear or identical and
returns zero for all negative values. The block diagram in Fig. 3 shows how the proposed model is used
for histopathological image classification and detection after training.

Fig. 2: Data Visualization

‘Larxel’ created a dataset of 10000 images related to colon adenocarcinoma and colon benign tissue
(COBT). There are 5000 images of colon adenocarcinoma and 5000 images of colon benign tissue among
them. The ‘Larxel’ dataset, which is available and taken from Kaggle LARXEL (accessed 2019), contains
images related to colon cancer. As illustrated in Fig. 2, the suggested research paper contains a total of
10,000 histopathological images.
After downloading the images, the dataset has been saved in a folder on our laptop and can be
accessed further in the next step. This research used a data of histopathological images in portable
network graphics (PNG) format. Colon adenocarcinoma and benign colon tissue were the two folders
into which the images were categorized.
During pre-processing, all of the images in a folder were read as input then resized before being used
in the proposed approach. Data is added once the loading is done and then converted to an array to
pre-process the image. Converted the images into the python numerical library (NumPy) which is used
to process values and speed up calculations. A label binarizer (LB) is used to distinguish integer values of
categories 0 and 1 to identify patterns. Then after labeling the images using the graphic tool ”Labeling”.
Label information for each image was generated for these labels in its datasets used for training as well
as testing purposes. The ratio of colon adenocarcinoma of 5000 images and 5000 images of colon benign
tissue data was determined by stratification. The 10,000 images of the total were used in the data, which
was splitted into training and testing groups at random. Two categories were used as the main divisions
of the collected images. A training dataset consists of 8000 images representing 80% of the whole dataset
is the first. The remaining 2000 images, i.e. 20% of the total data set, belong to the second category of
data called test data.
The base model is a pre-trained MobileNetV2 design imported from ImageNet. Five additional layers
for transfer learning are added to the base model, as shown in Table 1 of the modified MobileNetV2.
These layers include 7 × 7 average pooling, 256 dense of neural networks, 0.5 dropout size, and dense 1 of
8 B.Anil Kumar, Mohan Bansal

Fig. 3: Block Diagram

2. Using an average pooling layer and flattening to get the average values and allows to the combination
of the data into one dimension. These layers help improve the performance of the modified MobileNetV2
while at the same time with having lesser parameters. A fully connected layer, where each output depends
on each input, utilizes an activation function known as the rectified linear unit (ReLU) using a dense
layer. Two dense layers of binary classification added after a dropout layer help to prevent overfitting.
A number of output neurons is more than one, which utilizes the function of softmax. The base layer is
frozen so that previously acquired features are not lost.
The resulting dataset is used to train a new set of trainable layers, which are then used to identify
features that can be used to identify between colon adenocarcinoma and colon benign tissue. This model
is updated after saving the weights. This pre-trained method allows the model to secure the features
learned without losing them and benefit from biased weights without incurring additional computational
costs. The proposed approach is then described in summary. New training parameters for the modified
MobileNetV2 model include learning rates of 0.001 and batch sizes of 48 as well as 60 epochs. The Adam
optimizer is used in several applications of deep learning, including computer vision. Utilizing an Adam
optimizer, which adjusts both first as well as second phases of gradient that helped estimated learning
rate of each neural network weight.
As each batch is received, transfer learning applies an appropriate algorithm to change the weights of
the neural network. Further trained neural network then saved, helps to make predictions on the image
dataset.
A proposed model for colon cancer detection combines Caffe and a modified MobileNetV2, known
as the Caffe-MobileNetV2 (CMNV2) model. A CMNV2 model can be trained and used to distinguish
between colon adenocarcinoma and colon benign tissue. Later imported two prototxt files and used a
pre-trained model that required loading an improved MobileNetV2 model for classification, and loaded
a better-developed detector architecture such as the Caffe model.
Trained the DNN further using Caffe approach for detection and modified MobileNetV2 design for
classification, before obtaining the frame dimensions. A list of histopathological images, their associated
locations, and a list of predictions was initialized using data input. A MobileNetV2 base model was
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 9

inserted into his DNN model and processed along with the previously imported files. The model took
the loaded test images as input and transformed them into values for classification.
After Caffe-MobileNetV2 (CMNV2) design is loaded, imported test data of histopathological images
and predicted their features to identify colon cancer. Labels in the text include colon adenocarcinoma
and colon benign tissue detected. The red-green-blue (RGB) dimensions were used for image scaling.
The saved test data 20% or 2000 images are loaded for further processing in order to predict the
presence of colon cancer in two classes of histopathological images.
The output results of the predicted histopathological images are then shown.

5 Discussion of the proposed CMNV2 model‘s performance

‘Larxel’ created total of 10000 histopathological images, which are utilized in this research study, as
stated in Section 4. The dataset is split into 8000 images, which form 80% of it, and 2000 images, which
make up 20% of it. The suggested model is trained on 8000 images in total, and the model is tested
on 2000 images. 1000 images of colon adenocarcinoma and the remaining 1000 images of colon benign
tissue form up the test dataset. The CMNV2 proposed model‘s test accuracy for detecting colon cancer
is 99.95% after model training.

Fig. 4: (a) Normal confusion matrix (b) Suggested confusion matrix of CMNV2

The confusion matrix is presented as a summary utilizing a table to evaluate the classification effec-
tiveness of the methodology Room (2019). It is then further utilized for presenting evaluation metrics of
suggested CMNV2 technique. If a data point’s actual class and predicted class ‘colon adenocarcinoma’
are the same, it is said to be a true positive. If the data point actual class matches the expected class
of ‘colon benign tissue’, the result is a true negative. False positives occur when a data point’s actual
class ‘colon benign tissue’ does not match the predicted class ‘colon adenocarcinoma’. If a data point’s
actual class is ‘colon adenocarcinoma’, but the projected class is ‘colon benign tissue’, it is called a false
negative.
The confusion matrix for the classification of the standard and the proposed model is represented in
Fig. 4(a) and Fig. 4(b), respectively. Value of true positive (TP) is 1000, value of true negative (TN) is
999, value of false positive (FP) is 1, as well as value of false negative (FN) is 0. These values are shown
in Fig. 4(b).

6 Results and Discussion

The entire research was run on a laptop with a 64-bit operating system with an x64-based processor,
Intel(R) Xeon(R) W-2125 CPU @ 4.00GHz 4.01 GHz as well as 64.0 GB RAM. Many experimental
approaches in the proposed research were built and implemented using Jupyter Notebook tools and
10 B.Anil Kumar, Mohan Bansal

Python 3.9 kernel. This CMNV2 model developed is used for the classification and identification of
histopathological image datasets.

6.1 Accuracy

It is a measure of the overall correctness of a classification model. It represents the proportion of

instances that were correctly classified by the model out of the total number of instances which is given
in equation (1).

(T P + T N )
Accuracy = = 0.9995 = 99.95% (1)
(T P + F P + F N + T N )

6.2 Precision

calculates the ratio of events that were accurately predicted to those that were all positively projected.
Equation (2) gives the precision value for classification.
TP
P recision = = 0.9990 = 99.90% (2)
(T P + F P )

6.3 Recall

It is a measure of a classification model’s ability to correctly identify instances of a specific class, out
of all instances that actually belong to that class. Recall is also known as sensitivity, which is given in
equation (3).
TP
Recall = = 1 = 100% (3)
(T P + F N )

6.4 F1-score

It is a measure of a classification model’s overall performance that takes into account both precision
and recall. The F1-score is the harmonic mean of precision and recall, and it is a way to balance the
trade-off between these two metrics. It is given by the equation (4).

(2 × P recision × Recall)
F 1 − score = = 0.9995 = 99.95% (4)
(P recision + Recall)

6.5 Error rate

It is defined as the ratio of the total number of misclassifications to the total number of predictions
made by the model. It is given by the equation (5).

Errorrate = (1 − Accuracy) = 0.0005 = 0.05% (5)

According to the results, the proposed CMNV2 model uses lesser parameters than previous methods
and is more accurate in detecting colon adenocarcinoma and colon benign tissue.
CMNV2 proposed method makes predictions based on the patterns and labels in the training dataset.
Fig. 6 shows the classification results for the image data. Fig. 6a to Fig. 6h show the prediction of colon
adenocarcinoma. Fig. 6i to Fig. 6p show predictions for benign colon tissue. Our model has a 99.95%
accuracy in detecting upon its testing data. Fig. 7 shows curves of accuracy and loss of the training and
testing images. The training accuracy line in Fig. 7a is green and the test accuracy line is blue. The
training loss is represented by the green line in Fig. 7b and the test loss is represented by the blue line.
The accuracy obtained utilizing equation (1) is a good way of starting because the classes are evenly
distributed. As represented in Table 3 and displayed in Fig. 5(a), 5(b), 5(c), accuracy, precision, recall,
f1-score, as well as error rate of previous models and our CMNV2 model, showing different metrics are
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 11

Table 3: Comparison of the suggested CMNV2 model’s accuracy, precision, recall, f1-score and error
rate with the previous model’s

Sl.No. Model Name Precision (%) Recall (%) F1-score (%) Error rate (%)
1 MobileNetV2 80.89 78.90 69.02 18.74
2 ResNet50 86.02 83.80 88.20 10.20
3 ResNet101V2 86.89 88.09 83.47 9.55
4 DcnscNctl21 81.32 64 71.69 8.90
5 EfficientNetB0 81.45 87.90 73.50 8.20
6 VGG-16 88.80 89.33 90.02 7.10
7 DCNN 100 99.59 99.80 0.20
8 CMNV2 (Proposed Model) 99.90 100 99.95 0.05

compared Hasan et al. (2022). Our suggested CMNV2 design improves classification and detection of
colon adenocarcinoma as well as colon benign tissue in all metrics. Table 4 presents a summary of the
accuracy results of the proposed CMNV2 model and a comparison with other models. These results
demonstrate that the proposed new CMNV2 technique outperforms other existing models. Table 4 and
Fig. 8 show the accuracy of the CMNV2 model in accuracy. Compared to various other techniques, the
CMNV2 model requires lesser parameters to detect colon cancer Kumar and Bansal (2023).

Table 4: Different models accuracy comparison

Sl.No. Model Name Year Accuracy (%)

1 MobileNetV2 Hasan et al. (2022) 2022 81.26
2 ResNet101V2 Hasan et al. (2022) 2022 90.45
3 DenseNet169 and SVM Ohata et al. (2021) 2021 92.08
4 VGG-16 Hasan et al. (2022) 2022 92.90
5 ResNet-50 Bukhari et al. (2020) 2020 93.91
6 DL-DIP Masud et al. (2021) 2021 96.33
7 CNN Mangal et al. (2020) 2020 96.61
8 Transfer learning Hamida et al. (2021) 2021 99.12
9 Lightweight CNN Sakr et al. (2022) 2022 99.50
10 DarkNet-19 and SVM Toğaçar (2021) 2021 99.69
11 CNN based-MA ColonNET Yildirim and Cinar (2022) 2022 99.75
12 DCCN Hasan et al. (2022) 2022 99.80
13 CMNV2 (Proposed Model) 2023 99.95

7 Conclusion

The proposed CMNV2 technique provides a method to classify and detect colon cancer with a modified
MobileNetV2 method together with Caffe model. A transfer learning approach allowed the model to dis-
tinguish between two categories of colon adenocarcinoma and colon benign tissue. The image collection
is utilized to test the colon cancer classifier. Accuracy, precision, recall, f1-score, and error rate are used
to assess model efficiency. The dataset was successfully utilized using the proposed methodology. Detec-
tion of histopathological images in experimental results showed 99.95% accuracy. Our results indicated
that certain suggested new CMNV2 may play an important role in predicting the presence of colorectal
cancer and could really serve as a diagnostic platform of early detection as well as follow-up colorectal
cancer. The overall performance of suggested methodology is being analyzed utilizing a database having
histopathological images as well as the results are compared with other approaches that have worked
previously in this area. According to the results, our approach is more efficient than majority of deep
learning existing algorithms in detecting colorectal cancer. Our application helps pathologists look at
images of the colon to make the correct diagnosis when they want accuracy.
12 B.Anil Kumar, Mohan Bansal

(a) (b)

(c) (d)

Fig. 5: Comparison graphs of suggested CMNV2 and existing model’s accuracy, precision, recall,
f1-score and error rate

Author contributions: Conceptualization, Dr. Mohan Bansal; validation, formal analysis, research,
resources, writing review and editing, visualization, supervision and project management. Conceptual-
ization, B.Anil Kumar; methodology, software, writing, editing, original draught preparation. The final,
published version of the paper has been reviewed and approved by all authors.

Funding: No external funding is received for this research.

Institutional review: Not applicable.

Consent Informed: Not relevant.

Data availability: A part related to written coding of the proposed CMNV2 model and Predic-
tion results of a colorectal cancer diagnosis from histopathological images are made publicly avail-
able on the https://github.com/AnilKumargithu/COLON-cancer-Detection-of-CROHN-s-Disease/
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 13

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n) (o) (p)

Fig. 6: colorectal cancer detection of colon adenocarcinoma and colon benign tissue

tree/main open-source platform Github. Complete coding will be provided on request.

Conflicting interests: The authors disclose that they have no conflicting interests.
14 B.Anil Kumar, Mohan Bansal

Fig. 7: (a) Graph of training and testing accuracy (b) Graph of training and testing loss

Fig. 8: Different models accuracy graphs comparison

References

Mohammad Ali Abbas, Syed Usama Khalid Bukhari, Asmara Syed, and Syed Sajid Hussain Shah.
The histopathological diagnosis of adenocarcinoma & squamous cells carcinoma of lungs by artificial
intelligence: A comparative study of convolutional neural networks. MedRxiv, 2020.
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 15

Thaer SA Abdalla, Jan Meiners, Sabine Riethdorf, Alexandra König, Nathaniel Melling, Tobias Gorges,
Karl-F Karstens, Jakob R Izbicki, Klaus Pantel, and Matthias Reeh. Prognostic value of preoperative
circulating tumor cells counts in patients with uicc stage i-iv colorectal cancer. PLoS One, 16(6):
e0252897, 2021.
Mojtaba Akbari, Majid Mohrekesh, Shima Rafiei, S.M. Reza Soroushmehr, Nader Karimi, Shadrokh
Samavi, and Kayvan Najarian. Classification of informative frames in colonoscopy videos using con-
volutional neural networks with binarized weights. In 2018 40th Annual International Conference of
the IEEE Engineering in Medicine and Biology Society (EMBC), pages 65–68, 2018.
Sudabeh Alatab, Sadaf G Sepanlou, Kevin Ikuta, Homayoon Vahedi, Catherine Bisignano, Saeid Safiri,
Anahita Sadeghi, Molly R Nixon, Amir Abdoli, Hassan Abolhassani, et al. The global, regional, and
national burden of inflammatory bowel disease in 195 countries and territories, 1990–2017: a systematic
analysis for the global burden of disease study 2017. The Lancet gastroenterology & hepatology, 5(1):
17–30, 2020.
Marzieh Araghi, Isabelle Soerjomataram, Mark Jenkins, James Brierley, Eva Morris, Freddie Bray, and
Melina Arnold. Global trends in colorectal cancer mortality: projections to the year 2035. International
journal of cancer, 144(12):2992–3000, 2019.
Melina Arnold, Mónica S Sierra, Mathieu Laversanne, Isabelle Soerjomataram, Ahmedin Jemal, and
Freddie Bray. Global patterns and trends in colorectal cancer incidence and mortality. Gut, 66(4):
683–691, 2017.
Necmi Arslan, Ahmet Yilmaz, Ugur Firat, and M Tanriverdi. Analysis of cancer cases from dicle uni-
versity hospital; ten years experience. J Clin Anal Med, 9(2):102–106, 2018.
Tina Babu, Deepa Gupta, Tripty Singh, and Shahin Hameed. Colon cancer prediction on different
magnified colon biopsy images. In 2018 Tenth International Conference on Advanced Computing
(ICoAC), pages 277–280. IEEE, 2018.
Andrew A. Borkowski, Marilyn M. Bui, L. Brannon Thomas, Catherine P. Wilson, Lauren A. DeLand,
and Stephen M. Mastorides. Lung and colon cancer histopathological image dataset (lc25000), 2019.
Syed Usama Khalid Bukhari, Asmara Syed, Syed Khuzaima Arsalan Bokhari, Syed Shahzad Hussain,
Syed Umar Armaghan, and Syed Sajid Hussain Shah. The histological diagnosis of colonic adenocar-
cinoma by applying partial self supervised learning. MedRxiv, 2020.
Wojciech A Cieślikowski, Joanna Budna-Tukan, Monika Świerczewska, Agnieszka Ida, Michal Hrab,
Agnieszka Jankowiak, Martine Mazel, Michal Nowicki, Piotr Milecki, Klaus Pantel, et al. Circulating
tumor cells as a marker of disseminated disease in patients with newly diagnosed high-risk prostate
cancer. Cancers, 12(1):160, 2020.
Laura De Conti, Marco Baralle, and Emanuele Buratti. Neurodegeneration and rna-binding proteins.
Wiley Interdisciplinary Reviews: RNA, 8(2):e1394, 2017.
Satvik Garg and Somya Garg. Prediction of lung and colon cancer through analysis of histopathological
images by utilizing pre-trained cnn models with visualization of class activation and saliency maps. In
2020 3rd Artificial Intelligence and Cloud Computing Conference, pages 38–45, 2020.
A Ben Hamida, Maxime Devanne, Jonathan Weber, Caroline Truntzer, Valentin Derangère, François
Ghiringhelli, Germain Forestier, and Cédric Wemmert. Deep learning for colon cancer histopathological
images analysis. Computers in Biology and Medicine, 136:104730, 2021.
Mohamed Hammad, Abdullah M Iliyasu, Abdulhamit Subasi, Edmond SL Ho, and Ahmed A Abd
El-Latif. A multitier deep learning model for arrhythmia detection. IEEE Transactions on Instrumen-
tation and Measurement, 70:1–9, 2020.
Mohamed Hammad, Rajesh NVPS Kandala, Amira Abdelatey, Moloud Abdar, Mariam Zomorodi-
Moghadam, Ru San Tan, U Rajendra Acharya, Joanna Plawiak, Ryszard Tadeusiewicz, Vladimir
Makarenkov, et al. Automated detection of shockable ecg signals: A review. Information Sciences,
571:580–604, 2021.
Mohamed Hammad, Mohamed Bakrey, Ali Bakhiet, Ryszard Tadeusiewicz, Ahmed A Abd El-Latif, and
Pawel Plawiak. A novel end-to-end deep learning approach for cancer detection based on microscopic
medical images. Biocybernetics and Biomedical Engineering, 2022.
Imran Hasan, Shahin Ali, Habibur Rahman, and Khairul Islam. Automated detection and characteri-
zation of colon cancer with deep convolutional neural networks. Journal of Healthcare Engineering,
2022, 2022.
Alexander Hendricks, Greta-Lou Eggebrecht, Alexander Bernsmeier, Reinhild Geisen, Katharina Dall,
Anna Trauzold, Thomas Becker, Holger Kalthoff, Clemens Schafmayer, Christian Röder, et al. Iden-
16 B.Anil Kumar, Mohan Bansal

tifying patients with an unfavorable prognosis in early stages of colorectal carcinoma. Oncotarget, 9
(44):27423, 2018.
Alexander Hendricks, Katharina Dall, Burkhard Brandt, Reinhild Geisen, Christian Röder, Clemens
Schafmayer, Thomas Becker, Sebastian Hinz, and Susanne Sebens. Longitudinal analysis of circulating
tumor cells in colorectal cancer patients by a cytological and molecular approach: Feasibility and clinical
application. Frontiers in oncology, 11:2038, 2021.
Sebastian Hinz, Alexander Hendricks, Amke Wittig, Clemens Schafmayer, Jürgen Tepel, Holger Kalthoff,
Thomas Becker, and Christian Röder. Detection of circulating tumor cells with ck20 rt-pcr is an
independent negative prognostic marker in colon cancer patients–a prospective study. BMc cancer, 17
(1):1–11, 2017.
Yoon Suk Jung, Minkyung Han, Sohee Park, Won Ho Kim, and Jae Hee Cheon. Cancer risk in the early
stages of inflammatory bowel disease in korean patients: a nationwide population-based study. Journal
of Crohn’s and Colitis, 11(8):954–962, 2017.
B. Anil Kumar and Mohan Bansal. Face mask detection on photo and real-time video images using
caffe-mobilenetv2 transfer learning. Applied Sciences, 13(2), 2023.
Naresh Kumar, Manoj Sharma, Vijay Pal Singh, Charanjeet Madan, and Seema Mehandia. An empirical
study of handcrafted and dense feature extraction techniques for lung and colon cancer classification
from histopathological images. Biomedical Signal Processing and Control, 75:103596, 2022.
Min Seob Kwak, Hun Hee Lee, Jae Min Yang, Jae Myung Cha, Jung Won Jeon, Jin Young Yoon, and
Ha Il Kim. Deep convolutional neural network-based lymph node metastasis prediction for colon cancer
using histopathological images. Frontiers in Oncology, 10:619803, 2021.
LARXEL. Colon cancer histopathological images. https: // www. kaggle. com/ datasets/ andrewmvd/
lung-and-colon-cancer-histopathological-images , accessed 2019.
Jiaqi Lu, Ruiqing Liu, Yuejuan Zhang, Xianxiang Zhang, Longbo Zheng, Chao Zhang, Kaiming Zhang,
Shuai Li, and Yun Lu. Development and application of a detection platform for colorectal cancer
tumor sprouting pathological characteristics based on artificial intelligence. Intelligent Medicine, 2
(02):82–87, 2022.
Rumana Hamid Makhdoomi, Beenish Mukhtar Shah, et al. Pathology in colorectal malignancy. In New
Treatment Modalities in Rectal Cancer, pages 73–96. Springer, 2020.
Sanidhya Mangal, Aanchal Chaurasia, and Ayush Khajanchi. Convolution neural networks for diagnosing
colon and lung cancer histopathological images. arXiv preprint arXiv:2009.03878, 2020.
Mehedi Masud, Niloy Sikder, Abdullah-Al Nahid, Anupam Kumar Bairagi, and Mohammed A AlZain.
A machine learning approach to diagnosing lung and colon cancer using a deep learning-based classi-
fication framework. Sensors, 21(3):748, 2021.
Kiyoshi Masuda and Yuki Kuwano. Diverse roles of rna-binding proteins in cancer traits and their
implications in gastrointestinal cancers. Wiley Interdisciplinary Reviews: RNA, 10(3):e1520, 2019.
Elene Firmeza Ohata, João Victor Souza das Chagas, Gabriel Maia Bezerra, Mohammad Mehedi Hassan,
Victor Hugo Costa de Albuquerque, et al. A novel transfer learning approach for the classification of
histological images of colorectal cancer. The Journal of Supercomputing, 77(9):9494–9519, 2021.
Klaus Pantel and Catherine Alix-Panabières. Liquid biopsy and minimal residual diseaselatest advances
and implications for cure. Nature Reviews Clinical Oncology, 16(7):409–424, 2019.
F Perri, F Longo, M Giuliano, Francesco Sabbatino, Gennaro Favia, F Ionna, R Addeo, G Della Vittoria
Scarpati, G Di Lorenzo, and Salvatore Pisconti. Epigenetic control of gene expression: Potential
implications for cancer treatment. Critical reviews in oncology/hematology, 111:166–172, 2017.
Chat Room. Confusion matrix. Mach. Learn, 6:27, 2019.
Ahmed S Sakr, Naglaa F Soliman, Mehdhar S Al-Gaashani, Pawel Plawiak, Abdelhamied A Ateya, and
Mohamed Hammad. An efficient deep learning approach for colon cancer detection. Applied Sciences,
12(17):8450, 2022.
Luisa F Sanchez-Peralta, Luis Bote-Curiel, Artzai Picon, Francisco M Sanchez-Margallo, and J Blas
Pagador. Deep learning to find colorectal polyps in colonoscopy: A systematic literature review.
Artificial intelligence in medicine, 108:101923, 2020.
Paola Sena, Rita Fioresi, Francesco Faglioni, Lorena Losi, Giovanni Faglioni, and Luca Roncucci. Deep
learning techniques for detecting preneoplastic and neoplastic lesions in human colorectal histological
images. Oncology Letters, 18(6):6101–6107, 2019.
Rebecca L Siegel, Kimberly D Miller, and Ahmedin Jemal. Cancer statistics, 2019. CA: a cancer journal
for clinicians, 69(1):7–34, 2019.
Colon cancer detection for Crohn’s disease based on transfer learning approach with Caffe-MobileNetV2 17

Zarrin Tasnim, Sovon Chakraborty, FMJM Shamrat, Ali Newaz Chowdhury, H Alam Nuha, Asif Karim,
Sabrina Binte Zahir, Md Masum Billah, et al. Deep learning predictive model for colon cancer patient
using cnn-based classification. Int. J. Adv. Comput. Sci. Appl, 12, 2021.
Mesut Toğaçar. Disease type detection in lung and colon cancer images using the complement approach
of inefficient sets. Computers in Biology and Medicine, 137:104827, 2021.
Suat Toraman, Mustafa Girgin, Bilal Üstündağ, and İbrahim Türkoğlu. Classification of the likelihood
of colon cancer with machine learning techniques using ftir signals obtained from plasma. Turkish
Journal of Electrical Engineering and Computer Sciences, 27(3):1765–1779, 2019.
Who. “cancer”. https: // www. who. int/ news-room/ fact-sheets/ detail/ cancer , 2020.
Muhammed Yildirim and Ahmet Cinar. Classification with respect to colon adenocarcinoma and colon
benign tissue of colon histopathological images with a new cnn model: Ma colonnet. International
Journal of Imaging Systems and Technology, 32(1):155–162, 2022.
Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and
Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43–76, 2020.