Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання
The accurate pancreas segmentation process is essential in the early detection of pancreatic cancer. The pancreas is situated in the abdominal cavity of the human body. The abdominal cavity contains the pancreas, liver, spleen, kidney, and adrenal glands. Sharp and smooth detection of the pancreas f...
Збережено в:
| Дата: | 2022 |
|---|---|
| Автори: | , |
| Формат: | Стаття |
| Мова: | Англійська |
| Опубліковано: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2022
|
| Теми: | |
| Онлайн доступ: | https://journal.iasa.kpi.ua/article/view/265645 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | System research and information technologies |
| Завантажити файл: |  |
Репозитарії
System research and information technologies| _version_ | 1866302868164706304 |
|---|---|
| author | Kakarwal, Sangeeta Paithane, Pradip |
| author_facet | Kakarwal, Sangeeta Paithane, Pradip |
| author_sort | Kakarwal, Sangeeta |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2022-10-17T22:12:39Z |
| description | The accurate pancreas segmentation process is essential in the early detection of pancreatic cancer. The pancreas is situated in the abdominal cavity of the human body. The abdominal cavity contains the pancreas, liver, spleen, kidney, and adrenal glands. Sharp and smooth detection of the pancreas from this abdominal cavity is a challenging and tedious job in medical image investigation. Top-down approaches like Novel Modified K-means Fuzzy clustering algorithm (NMKFCM), Scale Invariant Feature Transform (SIFT), Kernel Density Estimator (KDE) algorithms were applied for pancreas segmentation in the early days. Recently, Bottom-up method has become popular for pancreas segmentation in medical image analysis and cancer diagnosis. LevelSet algorithm is used to detect the pancreas from the abdominal cavity. The deep learning, bottom-up approach performance is better than another. Deep Residual Network (ResNet-18) deep learning, bottom-up approach is used to detect accurate and sharp pancreas from CT scan medical images. 18 layers are used in the architecture of ResNet-18. The automatic pancreas and kidney segmentation is accurately extracted from CT scan images. The proposed method is applied to the medical CT scan images dataset of 82 patients. 699 images and 150 images with different angles are used for training and testing purposes, respectively. ResNet-18 attains a dice similarity index value up to 98.29±0.63, Jaccard Index value up to 96.63±01.25, Bfscore value up to 84.65±03.96. The validation accuracy of the proposed method is 97.01%, and the loss rate value achieves up to 0.0010. The class imbalance problem is solved by class weight and data augmentation. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2022.2.08 |
| first_indexed | 2025-07-17T10:27:58Z |
| format | Article |
| fulltext |
S.N. Kakarwal, P.M. Paithane, 2022
104 ISSN 1681–6048 System Research & Information Technologies, 2022, № 2
UDC 303.732.4, 519.816
DOI: 10.20535/SRIT.2308-8893.2022.2.08
AUTOMATIC PANCREAS SEGMENTATION USING
RESNET-18 DEEP LEARNING APPROACH
S.N. KAKARWAL, P.M. PAITHANE
Abstract. The accurate pancreas segmentation process is essential in the early detec-
tion of pancreatic cancer. The pancreas is situated in the abdominal cavity of the
human body. The abdominal cavity contains the pancreas, liver, spleen, kidney, and
adrenal glands. Sharp and smooth detection of the pancreas from this abdominal
cavity is a challenging and tedious job in medical image investigation. Top-down
approaches like Novel Modified K-means Fuzzy clustering algorithm (NMKFCM),
Scale Invariant Feature Transform (SIFT), Kernel Density Estimator (KDE) algo-
rithms were applied for pancreas segmentation in the early days. Recently, Bottom-up
method has become popular for pancreas segmentation in medical image analysis
and cancer diagnosis. LevelSet algorithm is used to detect the pancreas from the ab-
dominal cavity. The deep learning, bottom-up approach performance is better than
another. Deep Residual Network (ResNet-18) deep learning, bottom-up approach is
used to detect accurate and sharp pancreas from CT scan medical images. 18 layers
are used in the architecture of ResNet-18. The automatic pancreas and kidney
segmentation is accurately extracted from CT scan images. The proposed method is
applied to the medical CT scan images dataset of 82 patients. 699 images and 150
images with different angles are used for training and testing purposes, respectively.
ResNet-18 attains a dice similarity index value up to 98.29±0.63, Jaccard Index
value up to 96.63±01.25, Bfscore value up to 84.65±03.96. The validation accuracy
of the proposed method is 97.01%, and the loss rate value achieves up to 0.0010.
The class imbalance problem is solved by class weight and data augmentation.
Keywords: Deep Learning, Dice Coefficient, Fully Connected Layer (FCN), Resid-
ual Network (ResNet-18), Visual Geometry Group (VGG).
INTRODUCTION
Image splitting task is rigorously act as vital character in image investigation [1].
Image splitting task is beneficial for many applications like clinical image analy-
sis, disease detection of crop, traffic control observation, metallic surface crack
detection, and Aerospace image analysis. CT, MRI, PET, and supplementary im-
ages are used in the clinical image diagnosis and cure of illnesses. Pancreas seg-
mentation is challenging job in medical image investigation and analysis [2]. Ac-
curate organ segmentation and rapid processing are the major challenges in the
result of medical images. Computerized segmentation of several image subsec-
tions is useful to analyze anatomical organization as well as abdominal body.
Segmentation is played major role in visualization and diagnosis of clinical im-
ages. Sub-grouping is an important subject to several image-processing research.
Spleen, liver, kidney, pancreas is present in the abdominal CT images [3]. Bot-
tom-up approach and Top-down approach are applied to image splitting process
[4]. In Top-down approach, medical image segmentation is performed within
minimum time- period but less accuracy of segmentation. Bottom-up approach is
Automatic pancreas segmentation using ResNet-18 deep learning approach
Системні дослідження та інформаційні технології, 2022, № 2 105
efficient approach for medical image segmentation with high accuracy and mini-
mum time-period. Semantic segmentation is one approach of deep learning which
used for abdominal computed tomography. Deep learning model is popular ap-
proach of a machine learning methods. Deep learning model is dealing with algo-
rithms with hierarchical procedure layers [5]. It is experimenting nonuniform
transformations to view and gain data characteristics successfully [6]. Currently
Deep learning model is popular in various domains such as medical image analy-
sis, medical signal analysis, speech recognition, bio informatics, computer vision
[4]. Convolution neural networks, Generative Adversarial Networks, networks
with auto-encoder, and recurrent neural networks are prominent deep learning
approaches. These approaches are introduced and used in various task to map
with state-of-the-art results. In deep learning, network training is required with
dataset. For network training, set of convolution network, annotated dataset, op-
timizer, minibatch size, epoch, loss function is used. Dataset can be divided into
training, validation and testing purposed also. Fully convolution network (FCN)
was introduced by Long et al. [5]. In FCN, fully convolution layer is used as the
last fully convolution network layer. For more accuracy of dense pixelwise predi-
cation, the network is used fully convolution network. Semantic segmentation can
be performed by FCN. FCN architecture is built with pooling, upsampling and
convolution. FCN can perform predication of image within on single forward
pass [7].
MATERIALS AND METHODS
Image Dataset and Ground Truth Labeling
Bottom-UP approaches are applying on a dataset of 80 patient, 53 male and 57
female patients high resolution (512*512) CT scan images of 3D abdominal with
1,5–2,5 mm slice thickness range using Philips and Siemens MDCT scanners [2].
The 63 patients CT scans abdominal images are erratic recommend by an expert
radiologist from the Picture Archiving and Communications System (PACS). Na-
tional Institutes of Health Clinical Center is providing database of CT dicom im-
ages for abdominal,78 to 79 years patient age series with a mean of 46.8±16.7 are
used dataset [8]. DICOM medical images are converted to PNG image format.
The view of scans images was axial, sagittal and coronoal with 1,5 mm or 3 mm
thickness available [9]. Human expert labeling was performed under guideline of
certified radiologist. 3 labels are created for manual labeling like background,
pancreas, and kidney. Medical images are selected with different focal phase an-
gle of CT. Size and shape of pancreas is varies in different view of CT. Image
input size is 3255255 ,699 images are used for and 150 for testing.
LevelSets Algorithm
Levelset algorithm was presented by Osher and Sethian. It can use zero corre-
sponding exterior method [10]. Basic phenomena are to adjustment of portable
pathway of a two-dimensional arc into portable path of a 3D surface. In this, level
Set is actively participate toward indication of arcs as zero levelset of high dimen-
sion hyper-superficial [9]. Challenge of such exercise suggestions supplementary
faultless algebraic execution also operates topological variation without any prob-
S.N. Kakarwal, P.M. Paithane
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 106
lems [8]. The predictable set edge is described zero levelset of an implicit symbol
U of the evolving curve. In brief, evolve the implicit levelset function ),,( tyx
to represent evolution of curve )(t with a speed ),( yxF in normal direction.
Time t, zero levelset (x, y)/ 0),,( tyx defines evolved shape )(t :
0)),(( tt . (1)
Differentiate equation (1) with respect to t and apply chain rule,
tt
. , (2)
F denotes speed of curve in normal direction, describe below
FN
t
.
, (3)
where N denotes outwards normal, and elaborated by equation
N . (4)
Substitute Equation (4) and (3) into (2), and calculate evolution equation
for as
0 Ft .
This is fundamental equation of the levelset [6], and zero levelset represents
item shape curve:
}0),,(),({)( tyxyxt .
Function is generally determined grounded on signed distance calculation
to primary front. Relating to image field purely based on Euclidean distance be-
tween the curve and one image point. Describe below:
),(),( yxdyx ,
where ),( yxd — Euclidean distance from point to borderline. Sign: points inside
borderline (–) sign and outside (+) sign.
Evolution of borderline is elaborated by partial differential equation on zero
levelset of :
F
t
,
where F — known function, which calculated by the local curvature at the
zero levelset, i.e., )( FF , where :
2
322
22
)(
2
.
yx
xyxyyxyxx
.
Speed function F is represent as ),,( IGLFF , where L — local informa-
tion, which define by local geometric properties, G — global property of front
depends on contour and location of front; I — independent properties that inde-
pendent of front [11].
Automatic pancreas segmentation using ResNet-18 deep learning approach
Системні дослідження та інформаційні технології, 2022, № 2 107
The propagating task is characterized as:
||)(
GAI FFg
t
.
The term FA, causing front to consistently multiply or bond with a speed of
FA dependent on sign [12].
FG : portion that relate on geometry of front, such as own local curvature:
)),((1
1
),(
yxIG
yxgI
,
where IG — convolute image I with Gaussian smoothing filter G with
characteristic width of .
The form of speed function expressed as [9]: FGFAF . Via calculation
control objects or different image force which stated in common but simple form:
BFFAF GA )( ;
),(1(
1
frontimg CCdist
A
;
),( frontimg CCdistB .
dist() — appropriate distance function which is used to calculate variance of im-
age features between propagating front and propagated zones. B — additional
forces from image content. C — image content model, including any of the color,
shape, and texture features.
Automatic Seeded Region Growing Algorithms for LevelSet
General area developing algorithms determine pixel value physically from each
mark region as seed. The seed point normally contains of great uniform to neigh-
boring pixels and can be correspond to region [9]. The nature of seed point is very
close together to cluster center in clustering algorithm. AP clustering algorithm is
going to discover seed points in an image [13].
All data points as potential cluster centers, and computes accessibility and
accountability data by equations below iteratively between every two data points
to find characteristic centers of data points:
)},(),({max),(),(
,
kiSkiakiSkiR
kktk
,
if
kiiti
kirkkRkaiki
,.
},,0{max),(,0min,, ,
kiti
kiRkka
.
)),(,0(),( .
To receive the seed jx by equation below and cluster this holds ix :
)],(),([argmax
1
kiakiR
Nj
.
S.N. Kakarwal, P.M. Paithane
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 108
The simple linear iterative cluster (SLIC) algorithm is going for segmented
pixel-level image before using AP clustering process by combining adjacent pix-
els with similar characteristics to some irregular pixel blocks [14].
The region growing is computed by continually assimilation of seed points
also surrounding pixels and division of aim objects with background. Lastly de-
sign of ROIs is in rectangular boxes according to segmentation results [10].
Euclidean Distance runs to measure LAB color matches by equation
2
0
2
00
2
00 )()()( ojjj bbaaLLS .
PROPOSED METHOD
Residual Network (ResNet-18)
The training accuracy problem is solved by Residual Network-18 (ResNet-18).
)(xH is underlying mapping layer fit to new layer with x map inputs from previ-
ous layer. Residual function is stated that xxHxf )()( with same dimensions.
So updated residual function xxf )( is use in deep network layer [15].
Residual function is used to improve training accuracy value with mini-
mum time.
This residual function is used in every connected layer which formulated
using following expression:
xMxfF i }){,( . (5)
In this x and F are input and output dimensions, iM is layer used for
RELU. xxF )( is used as short connection in ResNet-18 (Fig. 1). Linear projec-
tion can be execute through sM and add into eq. (5) when dimension will match.
sM is square matrix which address the training accuracy problem when matching
dimensions condition satisfied [16]:
xMFF su .
n
Weight Layer
Weight Layer
f(x) Relu
F(x)+x
Relu
+
Fig. 1. ResNet-18 Relu Function
Automatic pancreas segmentation using ResNet-18 deep learning approach
Системні дослідження та інформаційні технології, 2022, № 2 109
T a b l e 1 . ResNet-18 model Network Layer Detail
Layer Name Output Size ResNet-18 Layer
Conv1 112×112 7×7.64 Stride,2
3×3 max pool, stride 2
Conv2 56×56 2
64,33
64,33
Conv3 28×28 2
128,33
128,33
Conv4 14×14 2
256,33
256,33
Conv5 7×7 2
512,33
512,33
1×1
Table 1 is depicted detail about convolution network layer used in ResNet-18
approach.
Data Augmentation
Class imbalance is problem in deep learning image segmentation. An image data
augmenter setup to an examine of pre-processing choices for image augmentation,
such as resizing, rotation and reflection. Random x and y direction translation is
performed by [10,10] and random rotation by [180,180]. Data augmentation is
beneficial to enhance output and results of deep learning approaches using set
novel and discrete examples to datasets [17]. Transformations in datasets by using
data augmentation approaches authorize companies to minimize these experimen-
tal costs. In the system, data augmentation is using reflection, translation, and ro-
tation to improve model prediction accuracy and reducing overfitting of data [18].
Loss Function
Cross Entropy. Cross -Entropy is used in ResNet-18 as loss function to relate
binary classification problems that calculate the probability of specific class or not
[19]. Let d and g represent the input image and related ground truth or manually
annotated image respectively. The key aim of segmentation is to learn relation of
d and g. Cross-entropy loss function CETL is shown as
N
i
iiiiCET tdfgtdfgL
1
)),(1(log)1()),((log ,
where i pixel index, N total pixel.
Dice Loss. To calculate overlap rate of predicated mask and ground-truth for
segmentation results Dice Score Coefficient is used. DSC is formulated as below:
gd
gd
gdDSC
)(2
),( .
S.N. Kakarwal, P.M. Paithane
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 110
Dice loss function can be formulated as per below expression:
)1(
c
cDSCL DSCL ,
where C is number of iteration.
Categorical Cross-Entropy Loss. It is used for multi-class classification
task. In this work 3 class are used liked background, pancreas, and kidney. This
model helps to detect object belong to which class among other classes.
Categorical loss function computed using following expression:
N
p
C
j
s
s
CCET
j
p
e
e
N
L log
1
,
where M is positive class, pse is score of positive class.
RESULTS
Evaluation Parameter
Global Accuracy (GA): GA is the ratio of perfectly segmented pixels divided by
the total number of pixels
ij
C
j
C
i
i
C
i
p
p
GA
00
0
,
Fig. 2. ResNet-18 Validation Accuracy
Automatic pancreas segmentation using ResNet-18 deep learning approach
Системні дослідження та інформаційні технології, 2022, № 2 111
where ijp is the number of pixels of class i segmented as belonging to class j .
Mean Accuracy (MA): Mean Accuracy is an next step of Global Accuracy,
in which the ratio of match pixels are computed in a per-class pattern and then
averaged over the total number of classes [19]:
ij
C
j
ii
C
i p
p
C
MA
1
1
00
.
Interaction of Union (IoU): Jaccard index alias IoU, it is correlation be-
tween segmented image (S) and annotated image (B). The value of IoU lies be-
tween 0 to 120.
gd
gd
gdJaccard
IoU, .
Mean Interaction of Union (MIoU): Mean IoU is calculated by the average
IoU value of all classes.
C
IoU
MIoU C
,
where total number of classes value is denoted by C. 3 classes are used like pan-
crea, kidney and background.
Weighted Interaction of Union:
C
pIoU
WIoU
i
C
iC
*
0 .
Weight value is number pixel in class. Average IoU of each class [19].
Bfscore. The bf score measure how close to the segmented boundary of an
input image matched with the annotated image. The BF score is calculated with
the help of the harmonic mean of the precision and recall values with a distance
error tolerance to decide whether a point on the segmented boundary has a corre-
lated to annotated boundary or not [20]:
precisionrecall
recallprecision
score
**2
.
Dice Coefficient. It is correlation between segmented image (S) and anno-
tated image (B). The value of dice coefficient lies between 0 to 1 and easily con-
verted into % for understanding purpose [20]:
)(
*2),(
BS
BS
BSdice
.
Sensitivity
TPFN
TP
ySensitivit
,
where TP is true positive, FN is false negative [20].
Specificity
TNFP
TN
ySpecificit
,
where TN is true negative, FP is false positive [20].
S.N. Kakarwal, P.M. Paithane
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 112
T a b l e 2 . ResNet-18 Model Compare with VGG-16 and VGG-19
Approach VA, % Bf, % Number of Layer Time MA, %
ResNet-18 97.01 84.06±3.96 18 242 min 42 sec 42.16
VGG-16 97.01 84.05±4.02 16 522 min 7 sec 42.439
VGG-19 96.90 56.62±2.48 19 632 min 44 sec 42.05
Table 2 is depicted performance of ResNet-18 as compared other deep learn-
ing approaches. Validation accuracy is more as compared to VGG-19. ResNet-18
is required less training time. The Bf score of ResNet-18 is higher as compared to
VGG-16 and VGG-19. ResNet-18 is using 18 number of layer where 17 number
are convolution layer used so accuracy is improved.
T a b l e 3 . ResNet-18 Model Compare with VGG-16 and VGG-19 using Evalu-
ation Parameter
Approach IOU Dice Sensitivity Precision
ResNet-18 96.63 01.25 98.29 00.63 96.78 00.03 96.64 00.12
ATLAS [19] 55.50 17.10 69.60 16.70 67.90 18.20 74.10 17.10
U-net Single Mode [20] 74.10 00.13 74.30 00.17 78.90 00.13
Frame-1[2] 57.2 25.40 68.80 25.60 72.50 27.20 71.50 30.00
Frame-2[2] 57.9 13.60 70.70 13.00 74.40 15.10 71.60 10.50
LevelSET[1] 28.18 14.07 42.26 16.37 98.48 00.03 98.70 00.03
Table 3 is depicted the performance of ResNet-18 deep learning approach as
compared to state-of-arts. ResNet-18 having more IoU value as compared to other
approach. In Dice index, ResNet-18 is achieved higher value (Fig. 3).
0
10
20
30
40
50
60
70
80
90
100
IOU D i ce S e n s i ti vi ty Pre ci s i o n
ResNet‐18 ATLAS [22] U‐net Single
Mode[22]
Frame ‐1[2] Frame ‐2[2] Leve lSET[1]
Fig. 3. ResNet-18 Result Comparison with State-of-Art
Automatic pancreas segmentation using ResNet-18 deep learning approach
Системні дослідження та інформаційні технології, 2022, № 2 113
Above figure is showing the ResNet-18 result in detail in terms of IOU,
Dice, Sensitivity and Precision (Fig. 4 and 5).
Above figure is consist of origan image (subfigure A, B, C, D, E, F, G), lev-
elset boundary image (subfigure A1, B1, C1, D1, E1, F1, G1), binary levelset
(subfigure A2, B2, C2, D2, E2, F2, G2), segmented image by levelset (subfigure
A3, B3, C3, D3, E3, F3, G3) and Segmented image using ResNet-18 (subfigure
A4, B4, C4, D4, E4, F4, G4).
Fig. 5. ResNet-18 Result Comparison with State-of-Art using Performance
Fig. 4. ResNet-18 Result Comparison with State-of-Art using Performance Parameter
S.N. Kakarwal, P.M. Paithane
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 114
CONCLUSION
ResNet-18 approach for automatic pancreas segmentation from CT scan images.
18 layers architecture is used for multiple organ segmentation from CT scan im-
age. In VGG-16 method, 16 convolution layers are used, in VGG-19, 19 convolu-
tion layers are used. In deep learning, training of network layer is time consuming
task. Levelset method is also used for pancreas segmentation but if seeded point is
wrong then output image is not correct. Levelset algorithm takes more time to
extract image with non-clear boundary of pancreas. Proposed method takes very
less time as compared to other deep learning method. Bfscore value of proposed
method is superior from VGG-19 method for same dataset. Sensitivity, Mean Ac-
curacy and Mean-bfscore value are superior to other methods. Dice coefficient,
jaccard Index are high as compared to state-of-art. Proposed method values as per
evaluation matrix parameters are high as compared to state-of-art. Loss function
value is less than other methods. In method, patch labeling is used so 64*64 patch
channel used so only one organ segmentation can be performed. In proposed
method, manually annotation process is performed by medical practitioner and
two classes is generated for pancreas and kidney. With the help of proposed
method many abdominal organ segmentations can be performed. Input image size
is 255*255*3 used so maximum pixel information is available in convolution
network. Noisy pixel image information can be easily omitted during dropout
network layer. Accurate pancreas shape and size is detected by proposed method,
but it fails to detect pancreas cancer affected areas in percentage. Pancreas size
and shape is available in 2D image using proposed method. 3D pancreas image
detection is future scope for proposed method.
AUTHOR CONTRIBUTIONS
Pradip Paithane: Conceptualization, Data curation, Methodology, Software,
Roles/Writing-original draft. Dr.S.N. Kakarwal: Formal analysis, Investigation,
Visualization, Project administration, Supervision, Validation, Writing - review
and editing.
REFERENCES
1. Pradip M. Paithane, Dr.S.N.Kakarwal, and Dr.D.V.Kurmude, “Top-Down Method
used for Pancreas Segmentation”, International Journal of Innovative and Exploring
Engineering (IJITEE), 9-3, pp. 2278–3075, 2020.
2. Amal Farag, Le Lu, Holger R. Roth, Jiamin Liu, Evrim Turkbey, and Ronald M.
Summers, “A Bottom-Up Approach for Pancreas Segmentation Using Cascaded Su-
perpixels and (Deep) Image Patch Labeling”, IEEE Transactions on image process-
ing, 26-1, 2017. doi:10.1109/TIP.2016.2624198.
3. Pradip M. Paithane and S.A. Kinariwal, “Automatic Determination Number of Clus-
ter for NMKFC-means algorithm on Image Segmentation”, IOSR-JCE, 17-1, 2015.
4. Pradip M. Paithane and Dr.S.N. Kakarwal, “Automatic Determination Number of
Cluster for Multi Kenel NMKFCM algorithm on image segmentation”, Intelligent
System Design and Applications, Springer Cham, 2017, pp. 80–89.
5. Justin Ker, Lipo Wang, Jai Rao, and Tchoyoson Lim, “Deep Learning Applications
in Medical Image Analysis”, IEEE Transactions, pp. 9375–9389, 2018.
Automatic pancreas segmentation using ResNet-18 deep learning approach
Системні дослідження та інформаційні технології, 2022, № 2 115
6. Yingege Qu, Pheng Ann Heng, and Tien-Tsin Wong, Image Segmentation using the
levelset Method. New York: Springer, 2004.
7. Xin-Jiang, Renjie-Zhang, and Shengdong-Nie, “Image Segmentation Based on Level
set Method”, International conference on Medical Physics and Biomedical Engi-
neering, Elsevier, 2012.
8. Joris R Rommelse, Hai-Xiang Lin, and Tony F. Chann, A Robust Level Set Algo-
rithm for Image Segmentation and its Parallel Implementation. Springer, 2014.
9. P.M. Paithane, S.N. Kakarwal, and D.V. Kurmude, “Automatic Seeded Region
Growing with Level Set Technique Used for Segmentation of Pancreas”, Proceed-
ings of the 12th International Conference on Soft Computing and Pattern Recogni-
tion (SoCPaR 2020), 1383.
10. Qianwen Li, Zhihua Wei, and Cairong Zhao, “Optimized Automatic Seeded Region
Growing Algorithm with Application to ROI Extraction”, IJIG, 17-4, 2017.
11. H.K. Abbas, A.H. Al-Saleh, H.J. Mohamad, and A.A. Al-Zuky, “New algorithms to
Enhanced Fused Images from Auto-Focus Images”, Baghdad Sci. J., 10, 18(1):0124,
2021
12. R.J. Mitlif and I.H. Hussein, “Ranking Function to Solve a Fuzzy Multiple Objective
Function”, Baghdad Sci. J., 10, 18(1):0144, 2021.
13. O. Bandyopadhyay, B. Chanda, and B.B. Bhattacharya, “Automatic Segmentation of
bones in X-ray images based on entropy measure”, Int. J. Image Graph., 16-01,
2016.
14. I. Bankman, Handbook of medical imaging: processing and analysis. New York:
Acdameic Press, 2000.
15. Shuo Cheng and Guohui Zhou, “Facial Expression Recognition Method Based on
Improved VGG Convolution Neural Network”, IJPRAI, vol. 34, no. 7, 2020.
16. Pikul Vejjanugraha, Kazunori Kotani, Waree Kongprawechnon, Toshiaki Kondo,
and Kanokvate Tungpimolrut, “Automatic Screening of Lung Diseases by 3D Active
Contour Method for Inhomogeneous Motion Estimation in CT Image Pairs”,
Walailak J. Sci.Tech, 18-2, 2021
17. Mizuho Nishio, Shunjiro Noguchi, and Koji Fujimoto, “Automatic Pancreas Seg-
mentation Using Coarse-Scaled 2D Model of Deep Learning: Usefulness of Data
Augmentation and Deep U-Net”, Appl. Sci., 2020. doi: 10.3390/app10103360.
18. Srikanth Tammina, “Transfer Learning using VGG-16 with deep convolution Neural
Network for classifying Images”, IJSR Publications, 9-10, 2019.
19. H.R. Roth et al., Deep Organ: Multi-level Deep Convolutional Networks for Auto-
mated Pancreas Segmentation. MICCAI, 2015.
20. Robin Wolz, Chengwen Chu, Kazunari Misawa, Michitaka Fujiwara, Kensaku Mori,
and Daniel Rueckert, Automated Abdominal Multi-Organ Segmentation With Sub-
ject-Specific Atlas Generation, IEEE Transactions On Medical Imaging, 32-9, 2013.
Received 04.01.2022
INFORMATION ON THE ARTICLE
S.N. Kakarwal, PES Engineering College, Aurangabad, MH, India.
Pradip M. Paithane, Dr. Babasaheb Ambedkar Marathwada University, Auran-
gabad, MH, India, e-mail: paithanepradip@gmail.com
АВТОМАТИЧНА СЕГМЕНТАЦІЯ ПІДШЛУНКОВОЇ ЗАЛОЗИ З
ВИКОРИСТАННЯМ RESNET-18 МЕТОДУ ГЛИБОКОГО НАВЧАННЯ /
С.Н. Какарвал, П.М. Паітане
Анотація. Точний процес сегментації підшлункової залози є важливим проце-
сом для раннього виявлення раку підшлункової залози. Підшлункова залоза
S.N. Kakarwal, P.M. Paithane
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 116
розташована в черевній порожнині тіла людини як і печінка, селезінка, нирки
та наднирники. Чітке та плавне виявлення підшлункової залози у черевній
порожнині є складною та виснажливою роботою у ході дослідження медично-
го зображення. Для сегментації підшлункової залози в перші дні застосову-
ються підходи «зверху-вниз», як-от новий модифікований алгоритм
кластеризації K-середніх (NMKFCM), масштабно інваріантне перетворення
ознак (SIFT), алгоритм оцінювання щільності ядра (KDE). Останнім часом по-
пулярний метод BottomUp для сегментації підшлункової залози в аналізі ме-
дичного зображення та діагностики раку. Алгоритм LevelSet використовується
для виокремлення підшлункової залози серед черевної порожнини. Поглибле-
не навчання, підхід «знизу-вгору» кращий, ніж інші. Глибока залишкова ме-
режа (ResNet-18) глибоке навчання, підхід «знизу-вгору» використовується
для виявлення точної та чіткої підшлункової залози за медичними зображен-
нями КТ. В архітектурі ResNet-18 застосовується 18 шарів. Автоматична
сегментація підшлункової залози та нирок виокремлюється із зображень КТ-
сканування із високою точністю. Запропонований метод застосовано на
медичній комп’ютерній томографії 82 пацієнтів. 699 зображень і 150 зобра-
жень із різними кутами застосовують для навчання та тестування відповідно.
ResNet-18 досягає значення індексу подібності кубиків до 98,29±0,63, значення
індексу Жакара до — 96,63±01,25, значення Bfscore — до 84,65±03,96.
Точність валідації запропонованого методу становить 97,01%, а значення
коефіцієнта втрат досягає 0,0010. Проблема дисбалансу класу вирішується за
допомогою ваги класу та збільшення даних.
Ключові слова: глибоке навчання, коефіцієнт кубиків, повністю підключений
шар (FCN), залишкова мережа (ResNet-18), група візуальної геометрії (VGG).
|
| id | journaliasakpiua-article-265645 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:27:58Z |
| publishDate | 2022 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/a5/db32c650e939dc024ea72e28869036a5.pdf |
| spelling | journaliasakpiua-article-2656452022-10-17T22:12:39Z Automatic pancreas segmentation using ResNet-18 deep learning approach Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання Kakarwal, Sangeeta Paithane, Pradip глибоке навчання коефіцієнт кубиків повністю підключений шар (FCN) залишкова мережа (ResNet-18) група візуальної геометрії (VGG) Deep Learning Dice Coefficient Fully Connected Layer (FCN) Residual Network (ResNet-18) Visual Geometry Group (VGG) The accurate pancreas segmentation process is essential in the early detection of pancreatic cancer. The pancreas is situated in the abdominal cavity of the human body. The abdominal cavity contains the pancreas, liver, spleen, kidney, and adrenal glands. Sharp and smooth detection of the pancreas from this abdominal cavity is a challenging and tedious job in medical image investigation. Top-down approaches like Novel Modified K-means Fuzzy clustering algorithm (NMKFCM), Scale Invariant Feature Transform (SIFT), Kernel Density Estimator (KDE) algorithms were applied for pancreas segmentation in the early days. Recently, Bottom-up method has become popular for pancreas segmentation in medical image analysis and cancer diagnosis. LevelSet algorithm is used to detect the pancreas from the abdominal cavity. The deep learning, bottom-up approach performance is better than another. Deep Residual Network (ResNet-18) deep learning, bottom-up approach is used to detect accurate and sharp pancreas from CT scan medical images. 18 layers are used in the architecture of ResNet-18. The automatic pancreas and kidney segmentation is accurately extracted from CT scan images. The proposed method is applied to the medical CT scan images dataset of 82 patients. 699 images and 150 images with different angles are used for training and testing purposes, respectively. ResNet-18 attains a dice similarity index value up to 98.29±0.63, Jaccard Index value up to 96.63±01.25, Bfscore value up to 84.65±03.96. The validation accuracy of the proposed method is 97.01%, and the loss rate value achieves up to 0.0010. The class imbalance problem is solved by class weight and data augmentation. Точний процес сегментації підшлункової залози є важливим процесом для раннього виявлення раку підшлункової залози. Підшлункова залоза розташована в черевній порожнині тіла людини як і печінка, селезінка, нирки та наднирники. Чітке та плавне виявлення підшлункової залози у черевній порожнині є складною та виснажливою роботою у ході дослідження медичного зображення. Для сегментації підшлункової залози в перші дні застосовуються підходи "зверху-вниз", як-от новий модифікований алгоритм кластеризації K-середніх (NMKFCM), масштабно інваріантне перетворення ознак (SIFT), алгоритм оцінювання щільності ядра (KDE). Останнім часом популярний метод BottomUp для сегментації підшлункової залози в аналізі медичного зображення та діагностики раку. Алгоритм LevelSet використовується для виокремлення підшлункової залози серед черевної порожнини. Поглиблене навчання, підхід "знизу-вгору" кращий, ніж інші. Глибока залишкова мережа (ResNet-18) глибоке навчання, підхід "знизу-вгору" використовується для виявлення точної та чіткої підшлункової залози за медичними зображеннями КТ. В архітектурі ResNet-18 застосовується 18 шарів. Автоматична сегментація підшлункової залози та нирок виокремлюється із зображень КТ-сканування із високою точністю. Запропонований метод застосовано на медичній комп’ютерній томографії 82 пацієнтів. 699 зображень і 150 зображень із різними кутами застосовують для навчання та тестування відповідно. ResNet-18 досягає значення індексу подібності кубиків до 98,29±0,63, значення індексу Жакара до — 96,63±01,25, значення Bfscore — до 84,65±03,96. Точність валідації запропонованого методу становить 97,01%, а значення коефіцієнта втрат досягає 0,0010. Проблема дисбалансу класу вирішується за допомогою ваги класу та збільшення даних. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2022-08-30 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/265645 10.20535/SRIT.2308-8893.2022.2.08 System research and information technologies; No. 2 (2022); 104-116 Системные исследования и информационные технологии; № 2 (2022); 104-116 Системні дослідження та інформаційні технології; № 2 (2022); 104-116 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/265645/261704 |
| spellingShingle | глибоке навчання коефіцієнт кубиків повністю підключений шар (FCN) залишкова мережа (ResNet-18) група візуальної геометрії (VGG) Kakarwal, Sangeeta Paithane, Pradip Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання |
| title | Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання |
| title_alt | Automatic pancreas segmentation using ResNet-18 deep learning approach |
| title_full | Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання |
| title_fullStr | Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання |
| title_full_unstemmed | Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання |
| title_short | Автоматична сегментація підшлункової залози з використанням ResNet-18 методу глибокого навчання |
| title_sort | автоматична сегментація підшлункової залози з використанням resnet-18 методу глибокого навчання |
| topic | глибоке навчання коефіцієнт кубиків повністю підключений шар (FCN) залишкова мережа (ResNet-18) група візуальної геометрії (VGG) |
| topic_facet | глибоке навчання коефіцієнт кубиків повністю підключений шар (FCN) залишкова мережа (ResNet-18) група візуальної геометрії (VGG) Deep Learning Dice Coefficient Fully Connected Layer (FCN) Residual Network (ResNet-18) Visual Geometry Group (VGG) |
| url | https://journal.iasa.kpi.ua/article/view/265645 |
| work_keys_str_mv | AT kakarwalsangeeta automaticpancreassegmentationusingresnet18deeplearningapproach AT paithanepradip automaticpancreassegmentationusingresnet18deeplearningapproach AT kakarwalsangeeta avtomatičnasegmentacíâpídšlunkovoízalozizvikoristannâmresnet18metoduglibokogonavčannâ AT paithanepradip avtomatičnasegmentacíâpídšlunkovoízalozizvikoristannâmresnet18metoduglibokogonavčannâ |