Deep Learning Classification on Early Detection of Pancreatic Cancer Using CT Scan
Bhagyashree Pramod Bendale
Research Scholar (CSE),
Department of Computer Science and Engineering,
MIT School of Computing,
MIT Art Design and Technology University,
Loni Kalbhor, Pune, Maharashtra, India.
bendale660@gmail.com
Dr. Sagar Tambe
Ph.D (CSE), Associate Professor,
Department of Computer Science and Engineering,
MIT School of Computing,
MIT Art Design and Technology University,
Loni Kalbhor, Pune, Maharashtra, India.
swatishirke@mituniversity.edu.in
Abstract
The
pancreas, a gland found beneath the stomach, is the site of a particular kind
of cancer called a pancreatic tumour. It is very difficult to identify the
pancreatic tumour. Therefore it is
necessary to employ a computer-aided diagnostic (CAD) system to identify
pancreatic tumour. In this paper, in-depth clinical data as well as accurate
assessment of images can be provided by artificial intelligence (AI) throughout
intervention. This effort designs a model for optimum deep learning based
pancreatic tumour as well as nontumor classification using CT pictures. The
suggested approach uses an (AWF) adaptive window filtering method to reduce
noise in order to identify and categorise the presence of pancreatic tumours
and nontumors. The image segmentation method uses which
involves dividing an image into meaningful and distinct regions or segments. Additionally, feature extraction using the UNET results in
an accumulation of feature vectors. For categorization reasons, pancreatic ductal adenocarcinoma (PDAC) is
predicted by pathological grade classification models for PDAC. A number of simulations
are conducted in order to confirm the PDAC technique's
enhanced efficiency, and the outcomes are examined from various angles. The
ODLPTNTC method showed excellent results compared to more current techniques,
according to a thorough comparative outcomes study.
Keywords- Computed Tomography
(CT), Adaptive Window Filtering (AWF), Cascade Forward Neural Network
(CFNN), Classification; Machine
learning, Feature Extraction; Deep Learning.
I.
Introduction
The pancreatic tumour, one of the most fatal types of cancer with stagnant
survival rates in the past few decades, has been declared irreversible [1]. Frequently,
advanced stages of pancreatic cancer are found, early detection is a
significant problem. A kind of cancer called the cells of the pancreas are
where pancreatic cancer develops, an organ found behind the stomach. One of the
most serious and most deadly forms of cancer is this one. Pancreatic cancer can
be of two primary categories such as Exocrine tumours are develop
from the exocrine glands that make digestive enzymes and are especially
prevalent. Adenocarcinoma is a particularly typical kind of exocrine pancreatic
cancer. Neuroendocrine tumors are arise from the pancreatic
hormone-producing cells. Due to the fact that pancreatic cancer may not
initially exhibit any symptoms, it is sometimes referred to as a
"silent" illness. However, if it gets worse, symptoms including
digestive issues, unexplained weight loss, changes in stool colour, and
jaundice (yellowing of the cheeks and eyes) might appear. The goal of ongoing
research is to enhance early detection techniques, create more efficient cures,
and comprehend the genetic and molecular causes of pancreatic cancer. Novel
medicines and strategies are being investigated in clinical trials to increase
survival rates. The categorization of pancreatic tumours using deep learning
algorithms and computer tomography (CT) images is a subject of continuing
investigation and growth in the fields of medical imaging and machine learning
[2]. It's vital to highlight that clinical and scientific uses of deep learning
algorithms for medical picture categorization, including pancreatic tumour
diagnosis, have showed potential. The effectiveness and efficiency of cancer
diagnosis can both be enhanced by these techniques. But meticulous validation,
regulatory permissions, and coordination with healthcare experts are necessary
for the use of such models in clinical practise. Currently, radiation treatment
with MRI monitoring is used to reduce tumour size, but anatomical alterations
like respiration are unaffected because of interpatient variability and
infarction [2]. The job of accurately and quickly identifying the pancreatic
tumour is difficult [3]. Early diagnosis, prompt treatment, and earlier
recognition are more crucial.
With the advancement of computer science and image processing technologies
for detection and diagnostics using computer-aided design (CAD) systems
became more technologically advanced. Radiation therapists are increasingly
using CAD systems to enhance diagnostic precision, help with disease
interpretation and detection, and lessen physician burden. Deep neural network
(DNN) technology was recently developed, expanding the need for health care.
Increased pathology in pancreatic cancer prompts significant focus on improving
efficient treatment and testing CAD systems when accurate pancreatic cancer
diagnosis is possible using segmentation. Therefore, it is necessary to create
a novel pancreatic segmentation mechanism. The segmentation of the pancreas in problems
with computed tomography (CT) continue that has not been overcome in this work.
CT is frequently used to diagnose and monitor PC patients. However, in up to 30% of
cases, a patient receives an incorrect or delayed PC diagnosis. Accurate focus
can be provided via image-guided therapy to enhance therapeutic possibilities. (AI)
Artificial intelligence has the potential to enhance as well as deliver precise picture
analysis for operational purposes and comprehensive diagnostic expertise [7].
In radiology, dermatology, and ophthalmology, image diagnosis tasks have
successfully benefited from recent improvements. Ophthalmology,
Dermatology and radiology, scanning detection tasks have all
benefited from recent developments. This work uses CT scan scan information
to create an optimum deep learning-based pancreatic tumour as
well as
nontumor classification. Adaptive window filtering (AWF) is a method that is
included to reduce any noise that may be present. Additionally,
feature extraction using the UNET results in an accumulation of feature
vectors. For
categorization reasons, pancreatic
ductal adenocarcinoma (PDAC) is predicted by pathological grade classification
models for PDAC. A number of simulations are conducted in order to confirm the PDAC technique's enhanced efficiency, and the outcomes are
examined from various angles. The ODLPTNTC method showed excellent results
compared to more current techniques, according to a thorough comparative
outcomes study.
The rest of the essay is prepared as
follows: We go into depth about the crisis response design process in the next
segment. We give an overview of the planning and creation of this particular
study in division 3. The experimental conclusions and Section 4 of this article
contains explanations on this research, It is followed by Section 5's
conclusion.
II.
Process for
design
A thorough analysis of the design process categorization models for
pancreatic tumors is provided in this section. Deep
learning is a family of techniques with enhanced performance for learning and
wide application horizons that employ numerous layers to gradually remove
characteristics from the source data [6]. In order to support disaster
management, deep learning approaches are especially well adapted to solve
problems involving, natural language processing, damage assessment, motion
detection, facial recognition and transportation prediction.
1. Data gathering: Collect a sizable
dataset of pancreatic tumours on CT scans, both benign and malignant
(cancerous) patients. Experienced radiologists should appropriately categorise
these images. Make sure the information is varied, containing
information on various pancreatic tumour sizes, forms, and locations.
2. Data
preprocessing: The data should be cleaned and prepared. To ensure that the pixel sizes
and values are uniform, normalise the photos. Increase the dataset's size and
variability by adding new data. This includes the ability to rotate, flip,
zoom, and change the brightness and contrast.
3. Division of Data: Produce test sets, validation and training sets from the dataset. 15% for testing
set, 15% for validation and 70% for training is a usual proportion.
4. Selecting a deep learning model: Select the best deep
learning architecture for classifying images. For picture examination, (CNNs) convolutional
neural networks are frequently employed.
5. Architectural models: Develop a framework
for the neural network system. Multiple layers of convolution, max-pooling
layers, and fully integrated layers are common components of a CNN architecture
for classifying images.
6. Model Education: Utilising the proper loss
coefficients and optimizers, the deep learning model to be trained through the
training data. To prevent overfitting, use strategies like early halting.
7. Validation and Hyperparameter Tuning: Validate the model's
performance using the validation dataset. Fine-tune the model and
hyperparameters as needed to improve performance.
8. Testing: Analyse the efficacy of the
completed model on the experimental set of data to get an accurate assessment
of its classification capabilities.
9. Interpretability and Visualization: Implement techniques
to make the model's decisions interpretable to medical professionals. This can
include techniques like Grad-CAM for highlighting the regions in the CT scan
that influenced the classification.his context.
10. Therapeutic Validation: Work together with
medical experts to confirm the model's correctness and use in a clinical
situation. This stage is essential to verifying that the model can support
diagnoses in the actual world.
11. Deployment: Integrate the model into clinical
workflows if it demonstrates to be accurate and trustworthy. This can entail
creating an intuitive user interface that radiologists can utilise in their
regular work.
12. Frequent Updating and Upkeep: To maintain the
model's performance current with changing medical knowledge and deep learning
technologies, ongoing monitoring and upgrades are required.
13. Ethics-Related Matters: Maintain compliance
with healthcare laws, such as HIPAA in the US, and protect the confidentiality
and safety of patient data.
14. Medical Research: To determine how
employing this technology may affect patient outcomes and healthcare practises,
consider conducting clinical trials.
15. Learning Efforts: Teach medical
practitioners how to use AI-assisted diagnostic tools so they can change their
workflow to accommodate the new technology.
III.
Design and
Development
1.
Obtaining information about CT scans: Data
on pancreatic patients is gathered from the (CPR) Central Person Registry and
the (DNPR),
Danish National Patient Registry utilising demographic data. Averaging
26.7 diagnostic codes per patient, DNPR covers roughly 229 million hospital
diagnoses for 8.6 million individuals.
Figure 1. Tissue from a pancreatic cancer CT scan.
2.
Adaptive Window Filtering: By applying a filter with
different parameters based on the properties of the picture, adaptive window
filtering is an approach used in picture processing as well as computer vision
to improve or filter pictures to remove noise [28]. The filter characteristics are modified using
adaptive window filtering in accordance with the local content of the picture.
This might enhance the quality of images.
Figure 1. Tissue from a pancreatic cancer CT scan.
3.
Segmentation
of Image: Segmentation of image
is a fundamental task in computer vision and image processing, which involves
dividing an picture into meaningful and distinct regions or segments. These segments
are often characterized by shared visual properties, such as color, intensity,
texture, or other features. This approach uses the preprocessed
picture as input during the image segmentation procedure to identify the
damaged areas in the CT image. Image segmentation has undergone a
revolution thanks to convolutional neural networks (CNNs). CNNs are used in
pixel-wise segmentation mask learning and prediction in architectures like
U-Net, FCN, and SegNet. They are excellent at many different segmentation tasks [27].
4. An Extraction of Features Approach Using the
UNet: 4 encoder units and 4 decoder units which
are joined by the U-shaped encoder-decoder network architecture, or UNET,
consists of a bridge.
Figure 2. UNet Architecture [2].
In order to extract pertinent
characteristics from input data, notably photographs, a UNet-based feature
extraction approach makes use of the UNet architecture. Its encoder-decoder
layout with skip connections makes it useful for a variety of computer vision
and image analysis work since it can capture fine-grained details and
multiscale information. Utilising the UNet structure for
obtaining pertinent characteristics from pictures or information is referred to
as a UNet-based feature extraction approach. Convolutional neural network (CNN)
structure known as UNet was initially created for segmentation of biomedical
images tasks but has subsequently found use in a number of different
disciplines. It is renowned for its capacity to preserve an overall context
while capturing minute details in a picture. In activities like medical image
analysis, computer vision, and remote sensing where high-resolution maps of
features are necessary, UNet is particularly well-liked [26].
5. Models for
classifying pathogenic grades in PDAC: The severity or aggressiveness of pancreatic ductal
adenocarcinoma (PDAC) is predicted by pathological grade classification models
for PDAC. The features of the tumour are revealed by pathological grading,
which is crucial information that can influence therapy choices and forecast
patient outcomes. The degree of differentiation and other histological
characteristics are frequently evaluated as part of the grading systems for
PDAC. Histological characteristics such tubular
development, nuclear pleomorphism, and mitotic activity are frequently used to
grade PDAC. The World Health Organisation (WHO) classification, which divides
tumours into three grades—(Grade 1) well-differentiated,
(Grade 2) moderately differentiated, and (Grade 3) poorly differentiated—is the most widely used
grading system for PDAC. Because of the
infrequent occurrence of clinical samples with extreme pathological separation
classes in the hospital, the scores with few specimens were merged in this
inquiry, and all specimens were assigned either of the two prediction labels:
low grade or high grade. Undifferentiated, poorly, and moderate-lowly
diferentiated pathologies were defined as high grade; moderately,
moderate-highly, and highly diferentiated pathologies were defined as low
grade. The segmentation results showed that the lesion
areas were removed from the 3D segmentation, CT and PET Mask data, respectively.
Figure 3. Classification using PDAC [3].
6. Procedure of
labelling PET/CT images: Labelling
PET-CT scans is a crucial step in many medical applications, including illness
detection, diagnosis, and therapy planning. Positron Emission Tomography and
Computed Tomography is referred to as PET-CT. Typically, the procedure entails
locating and labelling particular areas or structures on the pictures for
further study. A crucial stage in the
analysis of medical imaging, aiding research, diagnosis, and patient care, is
labelling PET-CT images. To ensure the reliability and clinical significance of
the labelled data, extensive cooperation between medical specialists,
annotators, and data scientists is required [29]. The analysis took into
account factors like the tumor's size, location, (SUVmax) standard uptake value, connection to the tissues
around it, normal liver parenchyma SUV mean, SUVR (tumor-tonormal liver standard uptake value ratio), existence
of distant metastases, the existence of metastasis to lymph nodes, including
observations obtained at various points in time.
Figure 4. Procedure of labelling PET/CT images [4].
7. Pathalogical
grading prediction model: In the
context of medical imaging and histology, a a model for predicting abnormal
grades is a machine learning or deep model for
learning created to identify the pathological grade
of a given sample, frequently a tissue biopsy or medical picture. In oncology,
where it is used to assess the seriousness and aggressiveness of the disease,
pathological grading is an required
component of disease diagnosis. Models for
pathological grading are useful tools that can help doctors diagnose patients
more precisely and consistently. To assure their clinical value and safety,
they should be created in close cooperation with domain specialists.
IV.
Experimental
Outcomes
We begin by gathering a big dataset of
individuals with benign and malignant (cancerous) pancreatic tumours using CT
scans. Using the benchmark BioGPS dataset from [5], the performance of
pancreatic tumour classification is examined in this section. Figure
5 displays a sample set of the CT scans that make up the dataset.
Figure 5. Pancreatic tumours images [7].
Table 1 provides a thorough review of
the proposed technique's comparative performance of categorization utilizing
various Training Sets.
Table 1: comparative performance of categorization utilizing
various Training Sets.
|
Variable |
Accuracy |
Sensitivity |
Recall |
Precision |
|
Training Set 40 |
76.55 |
98.88 |
76.55 |
85.23 |
|
Training Set 50 |
78.65 |
98.2 |
78.65 |
89.9 |
|
Training Set 60 |
96.46 |
98.1 |
85.23 |
90.17 |
|
Training Set 70 |
95.41 |
98.01 |
86.56 |
91.56 |
|
Training Set 80 |
98.1 |
97.23 |
89.9 |
98.88 |
|
Training Set 90 |
98.01 |
96.46 |
90.17 |
92.22 |
|
Training Set 100 |
92.22 |
96.24 |
91.56 |
92.45 |
|
Training Set 20 |
96.24 |
95.41 |
92.22 |
93.54 |
|
Training Set 110 |
98.2 |
95.22 |
92.45 |
95.22 |
|
Training Set 120 |
93.54 |
93.54 |
93.54 |
95.41 |
|
Training Set 130 |
97.23 |
92.45 |
95.22 |
96.46 |
|
Training Set 140 |
95.22 |
92.22 |
95.41 |
95.41 |
|
Training Set 150 |
92.45 |
91.56 |
96.24 |
98.1 |
|
Training Set 160 |
85.23 |
90.17 |
96.46 |
98.01 |
|
Training Set 170 |
89.9 |
89.9 |
97.23 |
92.22 |
|
Training Set 180 |
90.17 |
86.56 |
98.01 |
96.46 |
|
Training Set 190 |
91.56 |
85.23 |
98.1 |
96.24 |
|
Training Set 200 |
98.88 |
78.65 |
98.2 |
95.41 |
|
Training Set 210 |
86.56 |
76.55 |
98.88 |
95.22 |
|
Training Set 220 |
96.24 |
95.41 |
92.22 |
93.54 |
Figure
6: Comparison of different segmentation examples.
PDAC's pathological grade
classification mode's performance. In contrast to the clinical data model,
which had an AUC of The PET/CT system achieved an AUC of 0.99 in the training
group cohort, 0.72 in the cohort for validation, and 0.74 in the trial cohort,
compared to 0.95 in the training cohort, 0.68 in the validity group, and 0.68
in the trial cohort. To boost the model's efficiency and precision, we combined
the clinical approach alongside the PET/CT DL model to develop the
PET/CT+Clinical information model.
V.
Conclusion
In this work, a useful method is
developed to identify pancreatic tumours and nontumors and to categorise their
presence. This is the study that we are aware of that uses a DL method
to predict the preoperative pathological grade of PDAC using PET/CT. Combining
PET/CT characteristics with significant clinical data increased the model's
ability to predict outcomes. Using persistent real-world records of
illness courses and deep studying, we provide a method for forecasting the risk
of a rare occurrence but particularly dedicated malignancy. The
method we've suggested includes several steps, including preprocessing,
segmentation, feature extraction, classification, and parameter optimisation. Future
based
on DL segmentation approaches may be developed to enhance the technique's
classification performance.
References
[1]
J.
Behrmann, C. Etmann, T. Boskamp, R. Casadonte, J. Kriegsmann, and P. Maaβ,
“Deep learning for tumor classification in imaging mass spectrometry,”
Bioinformatics, vol. 34, no. 7, pp. 1215–1223, 2017
[2]
https://medium.com/analytics-vidhya/what-is-unet-157314c87634
[3]
https://www.mdpi.com/2072-6694/12/12/3656
[5]
W.
Xuan and G. You, “Detection and diagnosis of pancreatic tumor using deep
learning-based hierarchical convolutional neural network on the internet of
medical things platform,” Future Generation Computer Systems, vol. 111, pp.
132–142, 2020.
[6]
Design
of Optimal Deep Learning-Based Pancreatic Tumor and Nontumor Classification
Model Using Computed Tomography Scans Maha M. Althobaiti , 1 Ahmed Almulihi , 1
Amal Adnan Ashour , 2 Romany F. Mansour , 3 and Deepak Gupta,” Hindawi Journal
of Healthcare Engineering Volume 2022, Article ID 2872461, 15 pages https://doi.org/10.1155/2022/2872461”
[7]
F-FDG-PET/CT-based
deep learning model for fully automated prediction of pathological
grading for pancreatic ductal adenocarcinoma before surgery Gong
Zhang1,2†, Chengkai Bao3,6†, Yanzhe Liu2†, Zizheng Wang4†, Lei Du5 , Yue
Zhang3,6, Fei Wang2 , Baixuan Xu5*, S. Kevin Zhou3,6* and Rong Liu2*,” Zhang et al.
EJNMMI Research (2023) 13:49 https://doi.org/10.1186/s13550-023-00985-4”
[8]
T. G.
W. Boers, Y. Hu, E. Gibson et al., “Interactive 3D U-net for the segmentation
of the pancreas in computed tomog_raphy scans,” Physics in Medicine and
Biology, vol. 65, no. 6, p. 065002, Article ID 065002, 2020.
[9]
W.
Xuan and G. You, “Detection and diagnosis of pancreatic tumor using deep
learning-based hierarchical convolutional neural network on the internet of
medical things platform,” Future Generation Computer Systems, vol. 111, pp.
132–142, 2020.
[10]
H. Ma,
Z.-X. Liu, J.-J. Zhang et al., “Construction of a convolutional neural network
classifier developed by computed tomography images for pancreatic cancer
diagnosis,” World Journal of Gastroenterology, vol. 26, no. 34, pp. 5156–5168,
2020.
[11]
Y.
Luo, X. Chen, J. Chen et al., “Preoperative prediction of pancreatic
neuroendocrine neoplasms grading based on enhanced computed tomography imaging:
validation of deep learning with a convolutional neural network,”
Neuroendo_crinology, vol. 110, no. 5, pp. 338–350, 2020.
[12]
M. Fu,
W. Wu, X. Hong et al., “Hierarchical combinatorial deep learning architecture
for pancreas segmentation of medical computed tomography cancer images,” BMC
Systems Biology, vol. 12, no. 4, pp. 56–127, 2018.
[13]
K.
Men, X. Chen, Y. Zhang et al., “Deep deconvolutional neural network for target
segmentation of nasopharyngeal cancer in planning computed tomography images,”
Frontiers Oncology, vol. 7, p. 315, 2017.
[14]
L.
Shen, W. Zhao, and L. Xing, “Patient-specific reconstruction of volumetric
computed tomography images from a single projection view via deep learning,”
Nature biomedical engineering, vol. 3, no. 11, pp. 880–888, 2019.
[15]
K.
Dmitriev, A. E. Kaufman, A. A. Javed et al., “Classification of pancreatic
cysts in computed tomography images using a random forest and convolutional
neural network ensemble,” in Proceedings of the International Conference on
Medical Image Computing and Computer-Assisted Intervention, pp. 150–158,
Springer, Cham, December 2017.
[16]
K.
Manabe, Y. Asami, T. Yamada, and H. Sugimori, “Improvement in the convolutional
neural network for computed tomography images,” Applied Sciences, vol. 11, no.
4, p. 1505, 2021.
[17]
Y.
Ding, Q. Zhu, Z. Xing, and L. Li, “An adaptive-fuzzy filter algorithm for
vision preprocessing,” in Proceedings of the 2006 IEEE International Conference
on Robotics and Biomimetics, pp. 578–582, IEEE, Kunming, China, December 2006.
[18]
J. N.
Kapur, P. K. Sahoob, and A. K. C. Wongc, “A new method for gray-level picture
thresholding using the entropy of the histogram,” Computer Vision, Graphics,
and Image Processing, vol. 29, no. 3, pp. 273–285, 1985.
[19]
M. Li,
Y. Li, Y. Chen, and Y. Xu, “Batch recommendation of experts to questions in
community-based question-answering with a sailfish optimizer,” Expert Systems
with Applications, vol. 169, Article ID 114484, 2021.
[20]
S.
Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing between capsules,” in
Proceedings of the Proc. 31st Conf. Neural Inf. Process. Syst., pp. 1–11, Long
Beach, CA, USA, December 2017.
[21]
H.-C.
Li, W.-Y. Wang, L. Pan, W. Li, Q. Du, and R. Tao, “Robust capsule network based
on maximum correntropy criterion for hyperspectral image classification,” Ieee
Journal of Selected Topics in Applied Earth Observations and Remote Sensing,
vol. 13, pp. 738–751, 2020.
[22]
B.
Warsito, R. Santoso, H. Suparti, and H. Yasin, “Cascade forward neural network
for time series prediction,” Journal of Physics: Conference Series, vol. 1025,
no. 1, p. 012097, 2018.
[23]
A.
Fathy and H. Rezk, “Political optimizer based approach for estimating SOFC
optimal parameters for static and dynamic models,” Energy, vol. 238, p. 122031,
2022.
[24]
K.-L.
Liu, T. Wu, P.-T. Chen et al., “Deep learning to distinguish pancreatic cancer
tissue from non-cancerous pancreatic tissue: a retrospective study with
cross-racial external validation,” Be Lancet Digital Health, vol. 2, no. 6, pp.
e303–e313, 2020.
[25]
K. Si,
Y. Xue, X. Yu et al., “Fully end-to-end deep-learning based diagnosis of
pancreatic tumors,” Beranostics, vol. 11, no. 4, pp. 1982–1990, 2021.
