Advancements in Pancreatic Cancer Detection: A Comprehensive

Investigation of Convolutional Neural Network Applications

Wenhan Wang

Information Engineering, Shanghai University, Shangda Road, Shanghai, China

Keywords: Convolutional Neural Network, Pancreatic Cancer, Medical Image Analysis.

Abstract: Pancreatic cancer is one of the leading causes of cancer death and poses a major challenge to the current health

care system due to its early concealment and high mortality. At the same time, the excellent performance of

machine learning technology, especially convolutional neural network technology in the fields of medical

image target detection and semantic segmentation provides a new solution for the recognition and early

prevention of pancreatic cancer. This review introduces the application of Convolutional Neural Networks

(CNNs) in the field of pancreatic cancer detection in recent years, introduces the basic structure and function

of CNN model, and further introduces the characteristics of classical CNN models such as ResNet and

DenseNet and their applications in the field of pancreatic cancer detection. Three complex CNN models,

PANDA, YCNN and DACTransNet, are emphasized, and their structures, characteristics and applications in

pancreatic cancer detection are introduced. These models leverage CNN's ability to extract complex features

from medical images, facilitating precise tumor identification. Then, the user-friendliness and interpretability

of different models are discussed, and the lack of clinical evaluation in current studies is pointed out. Future

research may focus on improving the CNN architecture, enhancing model generalization, and addressing

interpretability issues to optimize clinical applications. This review provides insight into the current state and

prospects of CNN-based pancreatic cancer detection and outlines possible directions for future exploration.

1 INTRODUCTION

In the US, pancreatic cancer has a about 13% 5-year

survival rate, making it an extremely deadly illness.

(Siegel, 2024). By 2040, pancreatic cancer is

expected to overtake colorectal cancer becoming the

second-leading cause of cancer-related deaths in the

United States, where it presently ranks third.

(Halbrook, 2023). Smoking, obesity, type 2 diabetes,

and family history are associated with risk for

pancreatic cancer. Patients with locally located

tumours are frequently disregarded since they do not

exhibit any symptoms or have nebulous symptoms

until they become signs of advanced disease.

(Mizrahi, 2020). Computerized Tomography (CT) is

the primary imaging method used to detect and

evaluate pancreatic cancer, but the diagnostic

performance of this method depends on the

experience of the radiologist. In addition, in

pancreatic cancer detection, about 40% of tumors

smaller than 2cm evade CT detection (Kang, 2021),

https://orcid.org/0009-0008-6316-7031

and new methods are urgently needed to supplement

the judgment of radiologists to improve the sensitivity

and accuracy of pancreatic cancer detection.

In recent years, the use of computer-aided

detection with Artificial Intelligence (AI) has

demonstrated great potential in the medical field (Qiu,

2022; Shen, 2017). Medical imaging is considered an

important source of information needed to diagnose

diseases. Using a variety of ways to detect disease at

its earliest stages is one of the most important factors

in reducing cancer and tumor mortality. AI has

produced encouraging outcomes in categorization

and decision support. A kind of AI called machine

learning, or ML, has sped up a lot of medical research.

A kind of machine learning called "deep learning"

(DL) uses layers of neural networks to identify the

precise characteristics required for illness

identification (Plis, 2014; Tajbakhsh, 2016). Neural

networks are built using a collection of neurons

consisting of activation functions and parameters.

These neurons harvest and integrate information from

Wang, W.

Advancements in Pancreatic Cancer Detection: A Comprehensive Investigation of Convolutional Neural Network Applications.

DOI: 10.5220/0012961500004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 665-668

ISBN: 978-989-758-713-9

665

pictures to create a model that accurately represents

the complex connection that exists between diagnosis

and images.

One possible solution is to apply Convolutional

Neural Networks (CNNs) to the recognition of

pancreatic cancer as well. CNN is one of the main

algorithms of deep learning (DL) and is a type of

feedforward neural network with depth structure and

convolutional computation, which can effectively

extract features from image data and learn complex

data patterns, and is specially used for processing and

analyzing data with grid structure, especially image

and video data. In the identification of pancreatic

cancer, CT images of patients can be used as different

CNN models and the specific use of models can get

better detection results, but it is still difficult to

determine the best use of CNN and the future

development direction.

On the one hand, classical CNN models, such as

ResNet and DenseNet, have achieved certain effects

in the recognition of pancreatic cancer (Zavalsız,

2023; Huy, 2023). On the other hand, in order to

improve the recognition effect and reduce the amount

of computation, more complex models have been

applied (Kou, 2023; Cao, 2023; Dinesh, 2023). For

example, by stacking different models to create a

hybrid network architecture, or introducing other

algorithms outside the CNN model, these methods

have improved the recognition effect of pancreatic

cancer to a certain extent.

This review will briefly introduce the application

of CNN model in pancreatic cancer recognition, and

then provide the details of their implementation.

Then, the advantages and existing problems of these

recognition methods will be discussed. Finally, last

section will summarize this review and give the

conclusions drawn from the discussion.

2 METHOD

2.1 Framework of AI-Based Pancreatic

Cancer Detection

The construction process of AI-assisted pancreatic

cancer recognition system generally includes data set

establishment, data preprocessing, feature extraction,

model training, model evaluation, model testing and

other steps. Some relatively complete research studies

will carry out real-world deployment and clinical

testing.

In terms of the establishment of data sets, some

researchers choose to use existing public medical

image data sets, such as TCGA or TCIA, which are

large-scale public databases whose image modes

include MRI, CT, etc., for research use, while others

choose to cooperate with local medical institutions to

obtain de-identified patient data sets.

CNN extracts features from images by stacking

multiple convolution layers and pooling layers. The

pooling layer is utilised to shrink the feature map's

size, while the convolution layer is in charge of

collecting local features from the input. By stacking

multiple convolution layers and pooling layers, CNN

can gradually extract the abstract features of images.

The specific feature extraction is highly related to the

CNN structure selected by the fusion method, and the

feature extraction processes of different types of

model structures are very different. After feature

extraction, CNNs typically include one or more fully

connected layers that map features to the final

category score. Every neuron in the fully connected

layer is connected to every other neuron in the

preceding layer, and the activation function uses all

of the inputs from the previous layer to weight each

neuron's output.

Through the above operations, CNN will obtain

the expected output of the model, calculate the loss,

optimize the parameters, and conduct continuous

iterative training until the stopping condition is

reached. In general, the maximum number of

iterations and the loss threshold are used as stopping

conditions, but some models also use the results

evaluated by the validation set directly as criteria.

Validation sets are the subset of the data set dedicated

to evaluating the performance of the model, and in

addition to validation sets, there are test sets that

evaluate the generalization ability of the model by

identifying data that the model has not seen before.

2.2 Classical CNN Model for

Pancreatic Cancer Detection

Hoang Quang Huy et al. used DenseNet for cancer

detection. Multiple densely connected convolutional

layers make up a dense block, which is the

fundamental building block of DenseNet. Within the

dense block, the output of each layer is associated

with the output of all previous layers, and the feature

transfer is carried out by a dense connection, which

allows the model to better capture the details in the

pancreatic cancer image. In order to decrease the

dimension of the feature map, DenseNet further adds

a transition layer. This densely connected structure

enables the network to make full use of features, thus

improving the expressiveness of features and the

accuracy of the model (Huy, 2023).

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

666

Muhammed Talha Zavalsız et al. used ResNet for

cancer detection. The basic unit of ResNet is the

residual block, which contains two branches: the

main branch and the residual branch. The main

branch contains a series of convolutional layers,

while the residual branch adds the input to the output

by connecting across layers. This design allows the

network to learn residual mapping, helping to better

capture complex features in pancreatic cancer images.

At the same time, ResNet was pre-trained on a large

dataset, requiring only further training using

pancreatic cancer images (Zavalsız, 2023).

2.3 CNN Model with Hybrid Network

Architecture

2.3.1 PANDA

The deep learning model consists of a cascade of

three network phases, which increases the complexity

of the model and allows the model to accomplish

more complex and rich recognition tasks. The first

phase of the network used the nnU-Net model for

pancreas localization. In the second stage, CNNS

were constructed, and the classification heads were

used for lesion detection. The third stage is the

differential diagnosis of pancreatic lesions when

abnormal pancreatic lesions are found in the second

stage, and the characteristic prototype of pancreatic

lesions is automatically encoded by the auxiliary

memory transformer branch for more accurate fine-

grained classification (Cao, 2023).

2.3.2 YCNN

The deep learning model consists of a cascade of

three network phases, which increases the complexity

of the model and allows the model to accomplish

more complex and rich recognition tasks. The first

phase of the network used the nnU-Net model for

pancreas localization. In the second stage, CNNS

were constructed, and the classification heads were

used for lesion detection. The third stage is the

differential diagnosis of pancreatic lesions when

abnormal pancreatic lesions are found in the second

stage, and the characteristic prototype of pancreatic

lesions is automatically encoded by the auxiliary

memory transformer branch for more accurate fine-

grained classification (Cao, 2023).

2.3.3 DACTransNet

The two primary modules of DACTransNet comprise:

By combining the local features of CNN with the

global features of the VIT-based model—which

consists of alternating superimposed convolution

blocks and converter blocks—the hybrid CNN-

Transformer network, which serves as the model's

backbone, and the ASPP module with deformable

convolution improve the model's capacity to derive

distinctive features from medical images of

pancreatic cancer. With multiple expansions and in

multiple sensing fields, multiple filters and pooling

are used, the innovative deformable Porous Space

Pyramid (DC-ASPP) module used in this model

detects the characteristics of pancreatic cancer photos,

acquiring information on multi-scale irregular objects

(Kou, 2023).

3 DISCUSSION

First of all, the models based on CNN mentioned in

this paper have good performance in the task of

pancreatic cancer recognition and detection. However,

compared with the CNN model with complex

architecture, there is a certain gap in the

generalization of classical CNN model. In addition,

the recognition tasks of some models focus on the

binary classification of the presence or absence of

pancreatic cancer (Zavalsız, 2023; Huy, 2023), which

is not user-friendly and applicable, which is not

conducive to doctors to use the models for further

diagnosis, and does not fit the use scenario in the real

world, which brings great difficulties to clinical use.

On the contrary, the PANDA and DACTransNet

models can identify specific areas of pancreatic

lesions and generate visual results with good

interpretability. The PANDA model can directly

output the segmentation mask of detected masses and

the patient-level probability, providing more direct

interpretability (Kou, 2023; Cao, 2023).

In terms of the selection of data sets, some studies

chose public medical image data sets, such as TCGA

or TCIA, etc. (Zavalsız, 2023; Huy, 2023; Kou, 2023),

in which data bias, labeling standard deviation, data

potential bias and other problems will affect the

generalization performance of the model and make

the model obtain worse evaluation and test results.

Instead, some researchers have worked with

medical institutions to build custom datasets and even

conduct real-world clinical evaluations (Cao, 2023;

Dinesh, 2023). Among the model studies mentioned

in this paper, only Kai Cao et al. 's PANDA model

study conducted clinical validation and evaluation in

the real world, and verified the model through a large

number of tests (Cao, 2023). Through the diagnosis

of the same CT image with the clinician, the

reliability of the model decision-making can be

Advancements in Pancreatic Cancer Detection: A Comprehensive Investigation of Convolutional Neural Network Applications

667

strongly proved, and clinical use will accelerate the

process of its application in addition to verifying the

applicability of the model.

A large number of studies have proved the

superior performance of CNN-based models in the

identification of pancreatic cancer (Zavalsız, 2023;

Huy, 2023; Dinesh, 2023). Future studies should pay

more attention to the generalization ability of models

by building complex architecture CNN models, and

at the same time focus on the interpretability of

models. Studies of algorithms and models have

yielded excellent results in terms of performance, but

more research should be done to prove the

applicability and reliability of the models through

real-world clinical testing.

At the same time, considering the invisibility of

early pancreatic cancer and the high lethality of late

pancreatic cancer, binary classification of obvious

pancreatic cancer images cannot bring significant

improvement in the real world, and some models

should focus on more difficult to identify tasks, such

as pancreatic cancer classification and screening of

early pancreatic cancer, to provide AI assistance for

the prevention of pancreatic cancer. Some advanced

deep learning models widely used in other domains

(Sun, 2020; Wu, 2024) may be considered in the

future to improve the prediction performance for

pancreatic cancer.

4 CONCLUSIONS

This paper introduced the construction process of

CNN-based pancreatic cancer recognition system. In

addition, it also introduces the structure of some

classical CNN models, such as ResNet and DenseNet,

and briefly describe their applications in pancreatic

cancer recognition. In the same way, three complex

CNN models PANDA, YCNN and DACTransNet are

also mentioned, and how to design the structure of

these complex CNN models for pancreatic cancer

recognition is introduced. In the future, this kind of

research should pay more attention to how to identify

more subtle early pancreatic cancer images, enhance

the generalization ability of the model, especially in

the real-world clinical test, and improve the real-

world situation through research.

REFERENCES

Cao, K., Xia, Y., Yao, J., Han, X., Lambert, L., Zhang, T., ...

& Lu, J. 2023. Large-scale pancreatic cancer detection

via non-contrast CT and deep learning. Nature

medicine, 29(12), 3033-3043.

Dinesh, M. G., Bacanin, N., Askar, S. S., & Abouhawwash,

M. 2023. Diagnostic ability of deep learning in

detection of pancreatic tumour. Scientific Reports,

13(1), 9725.

Halbrook, C. J., Lyssiotis, C. A., di Magliano, M. P., &

Maitra, A. 2023. Pancreatic cancer: Advances and

challenges. Cell, 186(8), 1729-1754.

Huy, H. Q., Dat, N. T., Hiep, D. N., Tram, N. N., Vu, T. A.,

& Huong, P. T. V. 2023, March. Pancreatic Cancer

Detection Based on CT Images Using Deep Learning.

In The International Conference on Intelligent Systems

& Networks (pp. 66-72). Singapore: Springer Nature

Singapore.

Kang, J. D., Clarke, S. E., & Costa, A. F. 2021. Factors

associated with missed and misinterpreted cases of

pancreatic ductal adenocarcinoma. European

Radiology, 31, 2422-2432.

Kou, Y., Xia, C., Jiao, Y., Zhang, D., & Ge, R. 2023, July.

DACTransNet: A Hybrid CNN-Transformer Network

for Histopathological Image Classification of

Pancreatic Cancer. In CAAI International Conference

on Artificial Intelligence (pp. 422-434). Singapore:

Springer Nature Singapore.

Mizrahi, J. D., Surana, R., Valle, J. W., & Shroff, R. T. 2020.

Pancreatic cancer. The Lancet, 395(10242), 2008-2020.

Plis, S. M., Hjelm, D. R., Salakhutdinov, R., Allen, E. A.,

Bockholt, H. J., Long, J. D., ... & Calhoun, V. D. 2014.

Deep learning for neuroimaging: a validation study.

Frontiers in neuroscience, 8, 92071.

Qiu, Y., Wang, J., Jin, Z., Chen, H., Zhang, M., & Guo, L.

2022. Pose-guided matching based on deep learning for

assessing quality of action on rehabilitation

training. Biomedical Signal Processing and Control, 72,

103323.

Shen, D., Wu, G., & Suk, H. I. 2017. Deep learning in

medical image analysis. Annual review of biomedical

engineering, 19, 221-248.

Siegel, R. L., Giaquinto, A. N., & Jemal, A. 2024. Cancer

statistics, 2024. CA Cancer J Clin, 74(1), 12-49.

Sun, G., Zhan, T., Owusu, B.G., Daniel, A.M., Liu, G., &

Jiang, W. 2020. Revised reinforcement learning based

on anchor graph hashing for autonomous cell activation

in cloud-RANs. Future Generation Computer Systems,

104, 60-73.

Tajbakhsh, N., Shin, J. Y., Gurudu, S. R., Hurst, R. T.,

Kendall, C. B., Gotway, M. B., & Liang, J. 2016.

Convolutional neural networks for medical image

analysis: Full training or fine tuning?. IEEE

transactions on medical imaging, 35(5), 1299-1312.

Wu, Y., Jin, Z., Shi, C., Liang, P., & Zhan, T. 2024.

Research on the Application of Deep Learning-based

BERT Model in Sentiment Analysis. arXiv preprint

arXiv:2403.08217.

Zavalsız, M. T., Alhajj, S., Sailunaz, K., Özyer, T., & Alhajj,

R. 2023. A comparative study of different pre-trained

deeplearning models and custom CNN for pancreatic

tumor detection. International Arab Journal of

Information Technology.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

668