https://doi.org/10.56038/oprd.v3i1.

309

Conference Article

Comparative Analysis of Baseline Vnet and Unet

Architectures on Pancreas Segmentation
Azim Uslucuk1* , Hakan Öcal2

1 Intelligent Systems Engineering,

Graduate School, Bartın University, Bartın, Türkiye Orcid ID: https://orcid.org/ 0009-0000-0724-3337,
e-mail: [email protected] ,
2
Computer Engineering, Faculty of Engineering Architecture, and Design, Bartın University, Bartın,
Türkiye, Orcid ID: https://orcid.org/0000-0002-8061-8059, e-mail: [email protected],
* Correspondence: Azim Uslucuk; e-mail: [email protected]; Phone: +905423581256

(First received October 22, 2023 and in final form December 24 2023)

3rd International Conference on Design, Research and Development

(RDCONF 2023)
December 13 - 15, 2023

Reference: Uslucuk, A., Öcal, H. Comparative Analysis of Baseline Vnet and Unet Architectures
on Pancreas Segmentation. Orclever Proceedings of Research and Development,3(1), 146-157.

Abstract

The pancreas is one of the vital organs in the human body. It has an essential role in the
digestive system and endocrine system. Diseases such as cancer, diabetes, hormonal
problems, pancreatitis, and digestive problems occur in pancreatic disorders. In detecting
pancreatic disorders, first blood and urine tests are requested. If further examination is
needed, CT (Computed Tomography), MR (Magnetic Resonance), and EUS (Endoscopic
Ultrasonography) imaging methods are used. Pancreas segmentation is generally the
process of defining and drawing the lines of the pancreas from medical images such as
CT and MRI. The size and shape of the pancreas varies from person to person. Manual
segmentation of the pancreas is time-consuming and varies between physicians.
Recently, deep learning-based segmentation methods that achieve high-performance
results in organ segmentation have become trendy. In this study, Unet and Vnet
architectures were comparatively analyzed on the NIH-CT-82 dataset. As a result of the
ablation studies, a validation sensitivity of 0.9978 and a validation loss of 0.041 were
obtained in the Unet architecture. In the training with the Vnet architecture, 0.9975
validation sensitivity and 0.046 validation loss values were obtained, respectively.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 146

 https://doi.org/10.56038/oprd.v3i1.309

Keyw ords: Artificial Intelligence, Deep Learning, Artificial Neural Networks, Pancreas,
Segmentation.

1. Introduction

CT images are the first medical imaging used to diagnose pancreatic disorders. CT images
are taken quickly and, in some cases, allow images with higher contrast to be obtained
than MRI. Medical imaging is a non-invasive technique for examining internal organs
and is the most common form of examination used after laboratory tests [1], [2]. The
pancreas is one of the organs whose boundaries are difficult to determine due to its
irregular shape and dimensions that vary from person to person [3]. Accurate
segmentation of the pancreas, which occupies a small portion of CT images (<0.5%), is
critical for the diagnosis and treatment planning of Pancreatic Cancer, a highly fatal
disease [4], [5]. In the segmentation process of medical images, manual segmentation of
large amounts of images has been extensively studied by medical image reporting groups
due to concerns that it may cause distraction [6]. In order to prevent this situation, it is
essential to turn to computation-based systems. Although the concept of artificial
intelligence, first put forward in the 1950s, entered a winter period from time to time, its
development continued to accelerate after 2000. Deep learning, one of AI’s sub-branches,
is a prominent method in medical image processing. Studies in the field of image analysis
have shown that the success rates of segmentation with deep learning are high [7].
Computer systems that imitate the human brain’s learning style and processing logic are
called ANNs (artificial neural networks). Some ANNs used in pancreas segmentation are
as follows. CNNs (Convolutional neural networks) are used to find local features in
images. It generally defines the edges, textures, and shapes of images. CNN-based ANNs
have been proposed many times in the processing of medical images [8]. Densely
Connected Neural Networks(DNN) are used to process high-level features of images. U-
net is a network that generally combines the features of CNN and DNN [9]. Milletari et
al. proposed the Vnet network, which is a more flexible network for learning how to
process 3D images volumetrically [10]. In this study, Vnet and Unet architectures, which
are most used in segmentation, were comparatively analyzed.

In recent years, deep learning models consisting of convolutional neural network layers
have shown high segmentation performance [11]. The deep learning method, which
labels and makes inferences by identifying pixels of images of organs or lesions in
medical images, is significant in terms of its high success rate [12]. Derin et al. They
comparatively tested U-net and its different versions, Attention U-Net, Residual U-Net,

Online ISSN: 2980-020X https://journals.orclever.com/oprd 147

 https://doi.org/10.56038/oprd.v3i1.309

Attention Residual U-Net, and Residual U-Net++, using CT images from the NIH-CT82
dataset of 82 patients. They reported that the results of Residual U-Net stand out with the
highest score of 0.908 precision and 0.999 accuracy [3]. Wang et al. They proposed a dual-
input v-mesh fully convolutional network (FCN) to cope with the low texture contrast
that makes the segmentation task difficult. They reported that this method is more
performant than previous methods [4]. Paithane and Kakarwal used 12-layer deep
learning networks with four convolution layers in the LMNS-Net model for pancreas
segmentation and obtained a membrane similarity index score of 88.68 ± 57.49% [13]. In
a new deep learning method proposed for gross tumor volume segmentation from MRI
images (GTV) of pancreatic cancer patients, 126 image sets of 21 patients were used as
the data set. SegResNet, SegResNet 2D, and SwinUNETR were compared as DL
architectures. As a result, it was stated that training DICE = 0.88 and test DICE = 0.78
scores were obtained with the SwinUNETR model [14]. To reveal bottlenecks in pancreas
segmentation with deep learning, Zhang et al. Their review of ~51 articles examined
guidance and collaboration incentives to overcome the challenges of pancreas
segmentation algorithms [15].

2. Materials and Methods

This section will give comprehensive information about the data set and deep learning
architectures used for comparative analysis.

2.1. Preparing the dataset

The NIH-CT82 dataset consists of 82 3D CT abdominal contrast images obtained by the

National Institutes of Health Clinical Center from 53 male and 27 female subjects [16].
While 17 of the subjects have healthy kidneys, 65 of them have healthy pancreas. The ages
of the cases ranged from 18 to 76 years. CT images are 512x512 pixels in size. First, the
data set is split into 80% training and 20% validation. Various sizing and image
processing techniques were applied to the data set. Besides, images with the nifti format
were converted to .npy data formats to address computational limitations.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 148

 https://doi.org/10.56038/oprd.v3i1.309

Additionally, CT images were resized to 160x160x160. During resizing, the ROI region
was cropped and enlarged. A random flip along the axis was applied to the images.
Sample images from the dataset are shown in Figure 1.

Figure 1. Sample images from the dataset

2.2 Details of the Vnet and Unet Architectures

The Baseline Unet architecture used for comparative analysis is shown in Figure 2. As
seen in Figure 2, a Unet architecture ranging from 16 to 256 filters was used to segment
pancreas images. ReLU was used as a batch normalization and activation function in both
architectures to normalize the data. In the Unet architecture, 2x2x2 MaxPooling and
dropout value was used as 0.8. Double convolution was used in each layer of Unet. In
addition, in the Unet and Vnet architectures, 5x5x5 filters were used in the first and last
layers, while 2x2x2 convolutional filters were used in the other layers. The Vnet
architecture used in the study is shown in Figure 3. As seen in Figure 3, a Vnet
architecture ranging from 16 to 256 filters was used to segment pancreas images. Vnet
architecture, unlike Unet architecture, processes 3D images volumetrically. In addition,
Vnet differs from U-Net in that it uses the convolutions layer instead of the up-sampling
and down-sampling pooling layer. The idea behind V-Net is that using the Maxpooling
process causes much information loss, so replacing it with another series of convolution
operations without padding helps preserve more information.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 149

 https://doi.org/10.56038/oprd.v3i1.309

Figure 2. Baseline Unet Architecture

Figure 3. Baseline Vnet Architecture

Online ISSN: 2980-020X https://journals.orclever.com/oprd 150

 https://doi.org/10.56038/oprd.v3i1.309

2.4 Performance metrics

Accuracy metrics were used for comparative performance evaluations of the models. The
mathematical formulas of the metrics are shown in Equation 1. In the equations, FP means
Flase Positive, FN means False Negative, TP means True Positive, and TN means True
Negative.

𝑻𝑵+𝑻𝑷
ACC = 𝑻𝑵+𝑻𝑷+𝑭𝑵+𝑭𝑷 (1)

3. Result

This section gives comparative performance analyses of the employed methodolog ies
and comparative information about the models’ good and bad aspects.

3.1 Model’s implementation details

Training and validation of the recruited models were carried out on the NVidia RTX 4000
graphics card. For the models, batch_size is one, optimizer ADAM, and the learning rate
is 0.001 [17]. The proposed models were trained on the dataset for 250 epochs. The
algorithms of the models were implemented in the Anaconda ecosystem with the Python
3.8 programming language and Tensorflow_gpu 2.5 library. Sparse Categorical Cross
entropy loss was used to calculate the loss of the models.

3.2 Ablation studies

Considering the computational limitations in the inference studies, the models were
designed as U and V-shaped with 16 to 256 filters. It was observed that the training failed
when the model size was increased by further reducing the data set size. In the studies
conducted, dropout values for Unet were determined as 0.5 to 0.9, and 0.8 was selected
as the ideal value in the training. For Vnet, dropout values from 0.1 to 0.9 were applied,
and 0.2 was determined to be the perfect dropout value.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 151

 https://doi.org/10.56038/oprd.v3i1.309

3.3 Comparative performance analysis on the NIH-CT82 dataset

Figure 4 and Figure 5 show the training accuracy and training loss of the Baseline Unet
architecture. The training sensitivity of Baseline Unet varied between 99.78 and 99.76. The
educational loss of the hired architect started from 0.18 and decreased to 0.05. No
decrease in the training loss of the architecture was observed after 250 epochs.

Baseline Unet
0,998

0,9978

0,9976

0,9974

0,9972

0,997
Epoch 0

Epoch 32
Epoch 40

Epoch 72
Epoch 80
Epoch 88

Epoch 120
Epoch 128

Epoch 160
Epoch 168

Epoch 200
Epoch 208
Epoch 216

Epoch 248
Epoch 8
Epoch 16
Epoch 24

Epoch 48
Epoch 56
Epoch 64

Epoch 96
Epoch 104
Epoch 112

Epoch 136
Epoch 144
Epoch 152

Epoch 176
Epoch 184
Epoch 192

Epoch 224
Epoch 232
Epoch 240
Training Accuracy

Fig. 4. Baseline Unet Training Accuracy NHI-CT82 dataset

Online ISSN: 2980-020X https://journals.orclever.com/oprd 152

 https://doi.org/10.56038/oprd.v3i1.309

Baseline Unet
0,3

0,25

0,2

0,15

0,1

0,05

0
Epoch 0
Epoch 8

Epoch 40
Epoch 48
Epoch 56

Epoch 88
Epoch 96

Epoch 128
Epoch 136
Epoch 144

Epoch 176
Epoch 184
Epoch 192

Epoch 224
Epoch 232
Epoch 16
Epoch 24
Epoch 32

Epoch 64
Epoch 72
Epoch 80

Epoch 104
Epoch 112
Epoch 120

Epoch 152
Epoch 160
Epoch 168

Epoch 200
Epoch 208
Epoch 216

Epoch 240
Epoch 248
Training Loss

Fig. 5. Baseline Unet Training Loss NHI-CT82 dataset

Baseline Vnet architecture performed slightly better than Unet architecture despite the
high parameter values it used. Figure 6 and Figure 7 show the training accuracy and
training loss of the Baseline Vnet architecture, respectively. The designed model achieved
a performance value of 99.60 in training accuracy and a performance value of 0.036 in
training loss. One of the biggest reasons the Vnet architecture uses high parameters is the
convolution process instead of the Maxpooling layer used in the Unet architecture.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 153

 0,1
0,15
0,2
0,25

0
0,05
0,3
0,82
0,84
0,86
0,88
0,9
0,92
0,94
0,96
0,98
1
Epoch 0 Epoch 0
Epoch 8 Epoch 8

Online ISSN: 2980-020X

Epoch 16 Epoch 16
Epoch 24 Epoch 24
Epoch 32 Epoch 32
Epoch 40 Epoch 40
Epoch 48 Epoch 48
Epoch 56 Epoch 56
Epoch 64 Epoch 64
Epoch 72 Epoch 72
Epoch 80 Epoch 80
Epoch 88 Epoch 88
Epoch 96 Epoch 96
Epoch 104 Epoch 104
Epoch 112 Epoch 112
Epoch 120 Epoch 120
Epoch 128 Epoch 128
Epoch 136 Epoch 136

Training Loss
Baseline Vnet
Baseline Vnet

Training Accuracy

Epoch 144 Epoch 144

Epoch 152 Epoch 152

https://journals.orclever.com/oprd
Epoch 160 Epoch 160
Epoch 168 Epoch 168
https://doi.org/10.56038/oprd.v3i1.309

Epoch 176 Epoch 176

Epoch 184 Epoch 184
Epoch 192 Epoch 192
Epoch 200 Epoch 200

Fig. 7. Baseline Vnet Training Loss on NHI-CT82 dataset

Fig. 6. Baseline Vnet Training Accuracy on NHI-CT82 dataset

Epoch 208 Epoch 208

Epoch 216 Epoch 216
Epoch 224 Epoch 224
Epoch 232 Epoch 232
Epoch 240 Epoch 240
Epoch 248 Epoch 248

154
 https://doi.org/10.56038/oprd.v3i1.309

Comparative analysis results of Unet and Vnet architectures are shown in Table 1. As a
result of the analysis, Vnet, which has 22.1 million parameters, performed slightly better
than Unet, which has 2.94 million parameters. Due to the number of parameters, training
Vnet took almost four times longer.

Table 1. Comparative performance results of the models

Training Validation Epoch

Architectures Methods Parameters(M)
Accuracy(%) Accuracy(%)

Baseline Unet CT images 2.94 99.78 99.78

250
Baseline Vnet CT images 22.1 99.6 99.82

4. Discussion and Conclusion

In this study, Unet and Vnet architectures, which are most commonly used in the
literature with their various variations, were comparatively analyzed on NHI-CT82
Pancreas CT images. The analysis also showed that both models were robust in terms of
segmentation. However, as can be seen from the comparative analysis, Unet performed
the segmentation process with 7.5 times fewer parameters than Vnet. Additiona lly,
training for Vnet’s architecture took 4 four times longer than training for Unet’s
architecture. One of the main reasons for this situation is that the Vnet architecture, unlike
the Unet architecture, uses the Convolution process instead of the Maxpooling layer. The
main reason for using a convolution layer instead of Maxpooling in the Vnet architecture
is that Maxpooling may cause high-level features to be lost. However, as can be seen from
the performance results, these two types of architecture continue to be indispensable in
all kinds of segmentation tasks, especially the segmentation of medical images.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 155

 https://doi.org/10.56038/oprd.v3i1.309

References
[1] H. Kasban, M. A. M. El-Bendary, and D. H. Salama, “A Comparative Study of Medical
Imaging Techniques,” Int. J. Inf. Sci. Intell. Syst., vol. 4(2), pp. 37–58, 2015, [Online].
Available: https://ilearn.th-
deg.de/pluginfile.php/480243/mod_book/chapter/8248/updated_JXIJSIS2015.pdf
[2] M. Aljabri and M. AlGhamdi, “A review on the use of deep learning for medical images
segmentation,” Neurocomputing, vol. 506, pp. 311–335, Sep. 2022, doi:
10.1016/j.neucom.2022.07.070.
[3] A. Derin, C. Gurkan, A. Budak, and H. KARATAŞ, “Pancreas Segmentation Using U-Net
Based Segmentation Networks in CT Modality: A Comparative Analysis,” Avrupa Bilim Ve
Teknol. Derg., no. 40, pp. 94–98, 2022, Accessed: Oct. 07, 2023. [Online]. Available:
https://dergipark.org.tr/en/pub/ejosat/article/1171803
[4] Y. Wang et al., “Pancreas segmentation using a dual-input v-mesh network,” Med. Image
Anal., vol. 69, p. 101958, Apr. 2021, doi: 10.1016/j.media.2021.101958.
[5] Y. Zhou, L. Xie, W. Shen, Y. Wang, E. K. Fishman, and A. L. Yuille, “A Fixed-Point Model
for Pancreas Segmentation in Abdominal CT Scans,” in Medical Image Computing and
Computer Assisted Intervention − MICCAI 2017, M. Descoteaux, L. Maier-Hein, A. Franz, P.
Jannin, D. L. Collins, and S. Duchesne, Eds., in Bilgisayar Bilimleri Ders Notları. Cham:
Springer International Publishing, 2017, pp. 693–701. doi: 10.1007/978-3-319-66182-7_79.
[6] O. Oktay et al., “Attention U-Net: Learning Where to Look for the Pancreas.” arXiv, May 20,
2018. Accessed: Oct. 26, 2023. [Online]. Available: http://arxiv.org/abs/1804.03999
[7] T. ŞENTÜRK and F. LATİFOĞLU, “Biyomedikal Görüntülerin Bölütlenmesine Yönelik
Derin Öğrenmeye Dayalı Yöntemler: Bir Gözden Geçirme,” Dicle Üniversitesi Fen Bilim.
Enstitüsü Derg., vol. 12, no. 1, pp. 161–187, 2023, Accessed: Oct. 07, 2023. [Online]. Available:
https://dergipark.org.tr/en/pub/dufed/issue/75551/1181996
[8] F. Yuan, Z. Zhang, and Z. Fang, “An effective CNN and Transformer complementary
network for medical image segmentation,” Pattern Recognit., vol. 136, p. 109228, Apr. 2023,
doi: 10.1016/j.patcog.2022.109228.
[9] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical
Image Segmentation,” in Medical Image Computing and Computer-Assisted Intervention –
MICCAI 2015, vol. 9351, N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds., in
Lecture Notes in Computer Science, vol. 9351. , Cham: Springer International Publishing,
2015, pp. 234–241. doi: 10.1007/978-3-319-24574-4_28.
[10]F. Milletari, N. Navab, and S.-A. Ahmadi, “V-Net: Fully Convolutional Neural Networks for
Volumetric Medical Image Segmentation,” in 2016 Fourth International Conference on 3D
Vision (3DV), Oct. 2016, pp. 565–571. doi: 10.1109/3DV.2016.79.
[11]Ayhan M S, Kühlewein L, Aliyeva G, Inhoffen W, Ziemssen F, Berens P. “Expert-validated
estimation of diagnostic uncertainty for deep neural networks in diabetic retinopathy
detection”. Med. Image Anal. 64, 101724, 2020.

Online ISSN: 2980-020X https://journals.orclever.com/oprd 156

 https://doi.org/10.56038/oprd.v3i1.309

[12]A. G. Eker and N. Duru, “Medikal Görüntü İşlemede Derin Öğrenme Uygulamaları,” Acta
Infologica, vol. 5, no. 2, Art. no. 2, Dec. 2021, doi: 10.26650/acin.927561.
[13]P. Paithane and S. Kakarwal, “LMNS -Net: Lightweight Multiscale Novel Semantic-Net deep
learning approach used for automatic pancreas image segmentation in CT scan images,”
Expert Syst. Appl., vol. 234, p. 121064, Dec. 2023, doi: 10.1016/j.eswa.2023.121064.
[14]W. Choi et al., “Novel Deep Learning Segmentation Models for Accurate GTV and OAR
Segmentation in MR-Guided Adaptive Radiotherapy for Pancreatic Cancer Patients,” Int. J.
Radiat. Oncol., vol. 117, no. 2, Supplement, p. e462, Oct. 2023, doi:
10.1016/j.ijrobp.2023.06.1660.
[15]Z. Zhang, L. Yao, E. Keles, Y. Velichko, and U. Bagci, “Deep Learning Algorithms for
Pancreas Segmentation from Radiology Scans: A Review,” Adv. Clin. Radiol., vol. 5, no. 1,
pp. 31–52, Sep. 2023, doi: 10.1016/j.yacr.2023.05.001.
[16]H. R. Roth et al., “DeepOrgan: Multi-level Deep Convolutional Networks for Automated
Pancreas Segmentation.” arXiv, Jun. 21, 2015. Accessed: Dec. 05, 2023. [Online]. Available:
http://arxiv.org/abs/1506.06448
[17]D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” CoRR, Dec. 2014,
Accessed: Dec. 05, 2023. [Online]. Available:
https://www.semanticscholar.org/paper/a6cb366736791bcccc5c8639de5a8f9636bf87e8

Online ISSN: 2980-020X https://journals.orclever.com/oprd 157