U-Net: A valuable encoder-decoder architecture for liver tumors segmentation in CT images

Abstract

This study proposes a new predictive segmentation method for liver tumors detection using computed tomography (CT) liver images. In the medical imaging field, the exact localization of metastasis lesions after acquisition faces persistent problems both for diagnostic aid and treatment effectiveness. Therefore, the improvement in the diagnostic process is substantially crucial in order to increase the success chance of the management and the therapeutic follow-up. The proposed procedure highlights a computerized approach based on an encoder–decoder structure in order to provide volumetric analysis of pathologic tumors. Specifically, we developed an automatic algorithm for the liver tumors defect segmentation through the Seg-Net and U-Net architectures from metastasis CT images. In this study, we collected a dataset of 200 pathologically confirmed metastasis cancer cases. A total of 8,297 CT image slices of these cases were used developing and optimizing the proposed segmentation architecture. The model was trained and validated using 170 and 30 cases or 85% and 15% of the CT image data, respectively. Study results demonstrate the strength of the proposed approach that reveals the superlative segmentation performance as evaluated using following indices including F1-score = 0.9573, Recall = 0.9520, IOU = 0.9654, Binary cross entropy = 0.0032 and p-value <0.05, respectively. In comparison to state-of-the-art techniques, the proposed method yields a higher precision rate by specifying metastasis tumor position.

Keywords

Liver tumors CT images segmentation deep transfer learning encoder-decoder architecture

1 Introduction

In all over the world, liver cancer presents one of the foremost causes of death. The manual cancer tissue detection is a problematic aspect and time consuming. For accurate recognition and suitable therapy, a computer-aided diagnosis [1] is appreciated in decision making procedure. Accurate liver cancer detecting is consequently the chief aim using automated system.

Numerous medical imaging techniques such as ultrasonography (US), computed tomography (CT) and magnetic resonance imaging (MRI) are used for liver lesions examination. CT stills the vigorous imaging method to identify the cancerous tissue. Although, the qualified radiologists need to invasive methods, in some complicated cases, for confirming their diagnosis. In general, principal tumors for instance colon, pancreas and breast cancer frequently spread metastases to the liver in the anomaly development. Premature discovery of new-fangled liver metastases is still key because it can enhance the patient consequence. To explore the potential of computer-aided analysis, useful image processing techniques can be employed for liver tissue classification in order to assist the expert in diagnosis and decision-making method [2].

Artificial intelligence methods [3] have significance in several research applications for optimal characterization of the liver tumor. Regarding liver cancer detection, numerous systems are realized containing machine learning methods, watershed transform and region-based methods. A computer aided diagnosis (CAD) method proposed by Huang et al. [4] for segmenting and identifying the liver tumors using CT images. 81.7% of tumor classification precision is achieved in the extension work based on the auto-covariance texture features [5]. In clinical diagnosis of hepatocellular carcinoma, Zhou et al. [6] projected an important computational method focused on a particle swarm optimization technique. Newly, deep learning; in which neural networks and images are respectively processed of numerous layers and multiple types of filters; has been drawing attention as regards effective image pattern recognition [7 –12]. Deep convolutional neural network (CNN) method permits all of the information limited in the input image to be employed depending on the feature set picked out in convolutional machine learning. These methods [13 –15] achieve a pertinent performance in the pattern recognition of CT images which can successfully distinguish liver masses without radiologist intervention. In different computer vision domains such as region detection [16, 17], semantic segmentation [18] and image classification [19], deep learning schemes have reached massive achievement. Using CNN approach, alternating convolutional and pooling layers are provided to automatically extract multiple-level visual characteristics. This allows to obtain important progress in automated medical image diagnosis [20]. Melendez et al. [21] carried out many- instance learning focused on chest X-rays in order to detect the tuberculosis. It attained an AUC of 0.86. In the work of Hu et al. [22], a liver segmentation is appreciated based on CNN and globally optimized surface evolution, having 97% of mean Dice similarity coefficient (DSC). To classify skin cancer, Esteva et al. [23] trained a CNN and completed advanced accuracy compared to dermatologists. Besides, CNN has been applied in the segmentation process of different substructures for instance nuclei [24], neuronal structures [25], brain [26], cells [27], ventricles [28], and liver [29]. Still CNN presents the finest of used deep learning methods. In this context, Elshaer et al. [30] decreased the computation time of a great slices number based on two trained deep CNN models: The first one the liver region obtaining and the second for averting mistiness from image re-sampling and missed small lesions. In the work of Ma et al. [31], the CNN system is utilized in image patches; considering an image patch for each pixel (pixel of interest is in the patch. center). The patches are separated into two categories: normal or tumor liver. Least 50 percent or more of tumor tissue confirms that the patch is labeled as a positive subject. The authors report that the resulting accuracy reached 80.6%. ResNet, Alex Net, VGG-Net, etc. present different architectures of CNN model [32]. Additional researches [33, 34] have employed two-dimensional (2D) U-Net, intended mostly for medical image segmentation task.

In this work, two deep neural networks focused on U-Net and SegNet constructions [13, 14] were proposed to effectively resolve the image segmentation problems. In Table 1, an overview of previous approaches using CNN for different medical image segmentation applications. Using for increasing in attractiveness as relevant implementation of predicted segmentation learning systems, the U-net and Seg-net models present two important deep learning methods that have simpler models recognized as convolutional networks. This unsupervised learning offers a multilevel structure layer by-layer, automatically selecting increasingly more representations from the layers. The CNN Encoder-Decoder architecture mostly averts the gradient problem that can occur when training a standard neural network (without pre-initialisation). U-net pre-training increases the performance of the model by optimizing the CNN architecture and avoiding overfitting compared to others learning methods [14].

Table 1
Previous approaches using CNN in medical image segmentation problems

Applied CNN approaches Segmentation problems Modality Accuracy (%)

Standard CNN [35] Brain Breast Cardiac MRI- (T1, T2) 81.00

Volumetric U-net [36] Prostate Multiparametric-MRI 89.80

U-Net [37] Liver CT 95.05

3D-CNN [38] Brain MRI- T1, T2, Flair, and (DWI) 85.00

Saliency deep learning(SDL-Seg) [39] Breast cancer CT 76.40

Bayesian CNN [40] ventricle MRI-T1 92.00

Applied CNN approaches	Segmentation problems	Modality	Accuracy (%)
Standard CNN [35]	Brain Breast Cardiac	MRI- (T1, T2)	81.00
Volumetric U-net [36]	Prostate	Multiparametric-MRI	89.80
U-Net [37]	Liver	CT	95.05
3D-CNN [38]	Brain	MRI- T1, T2, Flair, and (DWI)	85.00
Saliency deep learning(SDL-Seg) [39]	Breast cancer	CT	76.40
Bayesian CNN [40]	ventricle	MRI-T1	92.00

Due to the deep learning methods need an important data base, these two architectures can supply accurate models even with moderately small datasets. An enhanced fully convolutional neural network (CNN) concept is obtained by the U-Net method [41] via CNN layers followed by up-sampling ones. In this paper, an automated segmentation is established using CT liver cancer images. The included lesions are in low-contrast heterogeneous medical volumes. The proposed contribution is divided into three folds. A fully convolutional CNN is firstly trained in order to prove the adaptability to tricky metastasis liver lesions segmentation. To automatically identify benign and malignant tumors of the liver images, this paper reports these outcomes and elaborates a CAD scheme. The proposed process is applied in order to compare the extracted data by means of their ability for differentiating between malignant and benign tissue. Performance analysis process is completed through means of accuracy and other measures using several experimentations.

This paper is organized as follows: Section 2 describes the CAD system, Section 3 deals with results and discussion of the algorithm and Section 4 concludes the work.

2 Material and methods

CAD model called watershed Gaussian based deep learning (WGDL) is proposed. This technique consists of intensity-based-segmentation to capably describe the tumor lesion in CT liver images. The proposed segmentation process is revealed in Fig. 1. In this paper, the proposed work is split into two important stages: the first step is introduced by testing some segmentation methods such as U-Net and SegNet. After that, the next step is affected to demonstrate the effective obtained results regarding to qualitative and quantitative results by computing the performance model.

Fig. 1

The proposed segmentation process.

2.1 Patients and lesions data: Overview

In this work, an overall of 400 images collected from diverse CT scan machine including 200 healthy and additional 200 malignant topics. The CT images considered were collected from the Rabta Hospital of Tunis. We have used the DICOM images from 2D axial slices. The exported data are in JPG format and 512×512 size. Total 200 patients with pathologically confirmed metastasis cancer were used to realize this current study. In this dataset, each patient has 1 tumor measured 2 cm or less; considered in stage 0 (very early). The tumor has not invaded the large blood vessels in the liver. About 500 slices for each patient are extracted. A total of 8,297 slices were used to train and validate the proposed model. In effect, we were selected the pertinent slices by removing the overlapping slices. The datasets affected by liver lesions are divided into two groups: 170 training and 30 testing datasets. In this study, the examined malignant tumor is hepatic metastases. An entire of liver metastases images were also collected from the CT imaging center of Charles Nicole Hospital. The patients’ median age was 54 years (between 30 and 79). An influential patient number were employed to directly train the proposed Encoder–Decoder CNN model.

2.2 Data augmentation

Data augmentation is carried out to the training dataset focused on random acts in terms of rotations, translations, scale and flips. Novel training examples were produced for each training epoch. Diverse examples were generated for 500 epochs in order to effectively train the deep neural network system. In fact, only linear transformations were employed for averting non-real cancers in the validation.

2.3 Hepatic metastases tumor segmentation phase

The cancer segmentation phase supplies two phases procedure as defined in [8]. It includes firstly an automated liver localization and secondly segmentation of metastasis tumor region. Due to CT tumor images are tissues external to the liver with an analogue intensity to the cancers, its segmentation is so difficult. The original input is simplified based on the prior location information and the likelihood intensity variety in liver region via histogram equalization. Similar region intensity to the liver refers the appearance of other organs in the simplified image can disturb the process of tumor segmentation. Actually, the deep learning techniques can resolve the difficulty of the medical image treatment. It leads to effectively ovoid the classical segmentation approaches such as mathematical morphologies, actives contours. The centroid of the largest connected region in the treated CT image is defined as the starting point for region development.

2.3.1 Encoder–Decoder Architecture

The Encoder–Decoder method remains a deep completely convolutional neural network structure [44]. The proposed architecture contains firstly an encoder and secondly a corresponding decoder networks attended by a pixel-wise classification process. Figure 2 shows the structure of the mentioned approach. In effect, the decoder layers are employed to delineate the low resolution from the encoder layers for a whole input resolution. This leads to improve a 2D mask integrating the tumor segmentation. The input image and the used mask are in the same size. In this paper, we propose an advanced method based on two dissimilar encoder–decoder structures. These architectures are Seg-Net [45] and U-Net [46].

Fig. 2

Encoder-Decoder structure.

2.3.2 Seg-Net method

Seg-Net technique was considered to well-organized CNN structure with regard to pixel-wise semantic segmentation. The used method is well known in different road scene understanding applications such as building, cars, and etc. Investigated by the CNN layers in VGG16 [47] structure, the encoder layers in Seg-Net architecture is employed. Hence, the decoder provides max-pooling indices stored and forwarded on behalf of the corresponding encoder’s layers. This technique achieves non-linear image upsampling of their input characteristic schemes. In the decoding procedure, the benefit of reemploying max-pooling indices is to enhance the process of boundary delineation. This method can be combined into any encoder–decoder structure having negligeable changesets. The sigmoid activation function is carried out in the last layer which can efficiency classify each pixel as a cancerous or background area. The applied Seg-Net architecture is exemplified in Fig. 3.

Fig. 3

Seg-Net architecture.

2.3.3 U-Net method

Suppling a rapid and exact 2D-3D image segmentation [48], U-Net [49] is an entirely CNN architecture. The main advantage of this theory raised in its forcefulness and efficiency results even when used small training data set. U-Net architecture is separated into two matches: primary down-sampling (SAMPLING) and additional up-sampling (UP-SAMPLING). As shown in Fig. 4, the upper part is a standard completely CNN. It includes a blocks sequence of 3×3 convolutions layers presented as follow:

A rectified linear unit (ReLU) activation function.

A max pooling with a 2×2 filter and two strides for down-sampling.

Fig. 4

The proposed U-Net architecture (AF: Activation Function, ReLU: rectified linear unit).

The feature number is doubled after each block. Regarding to the down part, we employed a 2×2 up-sampling. The feature number of maps is split into two after each block and enchained with the corresponding feature map of the left side which presented by:

A 3×3 convolution

A ReLU activation function.

In the last layer, a one-by-one convolution is appreciated using a sigmoid activation function. This later is applied for connecting each 32-component feature vector to the desired class. In our work, 26 convolutional layers along an entire number of 9, 239, 681 training parameters were achieved. The used U-Net method is revealed in Fig. 4.

2.4 Performance analysis

Different statistical measures can be used to control and evaluate the correctness and the effectiveness of the analysis results. In this study, the CNN architecture is appreciated to get a hopeful and truthful detection ratio. The segmentations accuracy is computed focused on the resemblance degree between the ground truth and the resulting outputs projected by the two applied architectures (U-Net and Seg-Net). Hence, the included evaluation measure shown by the mean IoU is exposed by the following equation: $IoU = \frac{TP}{TP + FP + FN}$ (1)

In our experiments, the evaluation of the proposed method is carried out using a statistical analysis in terms of Recall ‘Rec’, precision ‘PC’ and F1_score ‘F1’, of the hepatic metastasis’s tumor images. These criteria are frequently used for the evaluation of analytical tests: $Rec = \frac{VP}{VP + FN}$ (2) $PC = \frac{VP}{VP + FP}$ (3) $F 1 = 2 \times \frac{PC \times Rec}{PC + Rec}$ (4)

TP and TN: the true positive and true negative numbers, FP and FN: the false positive and false negative numbers, respectively.

Comparing actual performance in controlling tumor process, the one-sided Mann–Whitney U test is applied. This is a non-parametric test by verifying the null hypothesis. If the resulting segmentations and the ground truth derived from the same group, an important performance is achieved. In effect, a significant threshold is set to 0.05 which correspond to 95% of confidence for accepting or not the null hypothesis.

3 Results

In this study, we displayed an advanced experimentation to model hyper-parameter tuning based on two dissimilar structures of Encoder–Decoder architecture. Table 2 demonstrates the greatest significant parameters in terms of the obtained optimal values focused on the proposed Seg-Net and U-Net. A CNN (Seg-Net, U-Net) can vigorously segment hepatic metastases tumor lesion even if data augmentation placed in dissimilar orientations in order to have the property (invariance). Different experiments are realized to choose the optimum selected hyperparameters. In order to minimize the cost function, the stochastic gradient descent training method is applied to reduce the error rate between the actual and the desired output. In Table 3, all tested parameters are proved for final choice rationale.

Table 2
Optimum results according the used hyper-parameters via Seg-Net and U-Net models

Method Data augmentation Learning rate Momentum rate Batch size Weight initialization Training methods Learning rate decay Dropout rates

Seg-Net Linear 0.0005 0.99 4 Normal SGD Linear (0.000062) 0.30

U-Net Linear 0.0005 0.99 4 Normal SGD Linear (0.000062) 0.10

Method	Data augmentation	Learning rate	Momentum rate	Batch size	Weight initialization	Training methods	Learning rate decay	Dropout rates
Seg-Net	Linear	0.0005	0.99	4	Normal	SGD	Linear (0.000062)	0.30
U-Net	Linear	0.0005	0.99	4	Normal	SGD	Linear (0.000062)	0.10

Table 3

Tested hyper-parameters

Hyper-parameters	Data		Learning rate				Batch size					Momentum rate			Weight initialization			Training methods
Tests	L	nL	0.5	0.05	0.005	0.0005	2	4	8	16	32	0.8	0.9	0.99	N	U	GU	RMS	A	Ag	SGD

^*Linear “L”, No-linear “nL”, Normal “N”, uniform “U”, Glorot uniform “GU”, RMSprop “RMS”, Adam “A”, Adagrad “A”, Stochastic gradient descent “SGD”.

Training process was proved the highest data (85% of the database), the proposed techniques were validated on the remaining CT data (15% of the remaining database) by computing the precision score (mean IoU). The mean IoU (accuracy) attained 96.12% and 93.70% using U-Net and Seg-Net, respectively. Besides, F1-score, Recall and precision are respectively 95.73%, 95.20% and 96.52% for U-Net method compared to Seg-Net architecture 92.18%, 93.26% and 92.45%. The resulting loss error focused on binary cross entropy are 0.0032 and 0.048 for U-Net and Seg-Net, respectively. The segmentation results in term of qualitative evaluation of Seg-Net and U-Net methods are exhibited in Fig. 4. In this context, we report that U-Net segmentation is closer to the ground truth compared to the obtained results of Seg-Net. 200 CT images including cancer lesions are tested in the proposed liver tumor detection approach. The projected scheme splits into two phases: liver separation and tumor lesion detection. We applied the statistical analysis to effectively demonstrate the relevant difference between the predicted segmented region and the ground truth labels supplied by the two models Seg-Net and U-Net. Simulation results illustrated in Fig. 5 highlight the robustness of the proposed method using large CT images. The projected approach was able to estimate metastasis tumor segmentation with an interesting accuracy result in the total database. Using U-Net model, experimental results are analyzed by computing the different statistical measures: the mean IoU, Recall ‘Rec’, precision ‘PC’ and F1_score ‘F1’. From the obtained results in Fig. 5, the performance of a segmentation algorithm usually depends on sensitivity and robustness; corresponding to the highest value of these statistical measure.

Fig. 5

The obtained results in view of U-Net and Seg-Net architectures.

The p-values are computed between the ground truth and the two predicted segmentations. After several experiments, we can notice that is no important difference between the U-Net segmentation and the ground truth one which reaches a p-value of 0.035. Compared to Seg-Net architecture, a failed result is attained (p-value = 0.057) in order to determine the difference between the ground truth labels and the Seg-Net’s segmentation. The two target segmentations furnished by U-Net and Seg-Net methods are different depending on U-test. For the validation results, we applied a statistical analysis using Student’s t–test in order to evaluate the difference between the predicted segmentation provided by the two architectures and the ground truth. In fact, if, p-value is inferior to 0.05 (p-value <0.05) then, it was considered significant. Besides, for each experiment, the pertinent binary cross entropy is depended on the lower concerned value. Figure 5 shows the obtained results using the U-Net and Seg-Net methods.

In this paper, we introduced an advanced approach based on two different fully CNN encoder–decoder architectures. These U-Net and Seg-Net structures were effectively realized to accomplish hepatic metastases tumor segmentation. As shown in Fig. 6, the U-Net segmentation results displayed better accuracy compared to the manual segmentation of the experts. It is clear that Seg-Net segmentation results are not highly near to the ground truth. In fact, the foremost variation between the two applied models is presented in their connection type in view the convolution and deconvolution. Seg-Net employed merely the kept pooling indices through convolution process compared to U-Net is focused on features maps linking in terms of convolution and deconvolution. In this work, the attendance of tumor lesion can be exposed in rapports of analysis reliability of the hepatic metastasis’s cancer localization by different evaluations. Three radiologists, with varied expertise degrees, contributed in specifying the liver tumor ground truth, via the attentive lesion detection on each patient.

Fig. 6

Quantitative segmentation results using U-Net and Seg-Net architectures.

4 Discussion

The majority of current approaches used segmentation improvement techniques based on statistical form and emergence models. In the work of Lu et al. [50], authors proposed an advanced method focused on multiple features for training the kernel SVM via RBF kernel. This method can vigorously distinguish the hepatic lesion and the surrounding tissues through their corresponding statistical features. Still, this study requires an enhancement characterization stage by reducing the feature number. Besides, an important approach for metastases liver detection is applied in the work of Albishri et al. [51] based on multimodal U-net model. The authors notice that the segmentation results are efficiency improved than those obtained by atlas-based methods. Still, deep learning methods confirm the high potential in the medical segmentation applications. But the learned features require to be trained on a greater data base. For an improved qualitative analysis, a statistical measurement could be used to compute the quality of image segmentation. Table 4 reveals that the segmentation accuracy of the proposed method demonstrates operative segmentation precision using our dataset with a dice average of 93.54%, compared to other tested state-of-the-art techniques [50] and [51]. Finally, the proposed CT image segmentation approach can faithfully be exploited for the analysis of metastases tumor in precocious stage and used consequently as a further tool to assist radiologist in the diagnosis process.

Table 4
Summary of some studies reporting some liver lesion segmentation approaches.

Literature Year Used techniques Used Dataset Dice (%)

Lu et al. [50] 2012 SVM CT 85.38

Albishri et al. [51] 2019 U-net segmentation CT 89.64

Proposed method U-Net CT 93.54

Literature	Year	Used techniques	Used Dataset	Dice (%)
Lu et al. [50]	2012	SVM	CT	85.38
Albishri et al. [51]	2019	U-net segmentation	CT	89.64
Proposed method		U-Net	CT	93.54

5 Conclusion

In this study, we proposed and tested a new automated tumor segmentation approach based on two different encoder–decoder architectures: SegNet and U-Net. In each architecture, all layers and parameters are rectified in order to repeat the deep neural network training based on computed tomography data. The U-Net model highlights the finest segmentation results. To enhance the forcefulness of the proposed model, a novel data can be added that is still following annotation. We plan to use the segmented cancer approach for advanced tumor stages to efficiency improve prediction system and provide a prospective validation. Besides, the proposed segmented hepatic metastases method can be deployed to other lesion organs such as breast cancer response.

References

Chang

C.C.

, Chen

H.H.

, Chang

Y.C.

, et al., Computer-aided diagnosis of liver tumors on computed tomography images, Computer Methods and Programs in Biomedicine 145 (2017), 45–51.

Anwar

S.M.

, Awan

, Yousaf

, et al., Segmentation of liver tumor for computer aided diagnosis. In 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES) (2018), 366–370.

Nishida

and Kudo

, Artificial intelligence in medical imaging and its application in sonography for the management of liver tumor, Frontiers in Oncology 10 (2020), 2862.

Huang

, Yang

, Lin

, et al., Random feature subspace ensemble based extreme learning machine for liver tumor detection and segmentation. In 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (2014), 4675–4678.

Zhou

, Shi

, Zhu

, et al., Shearlet-based texture feature extraction for classification of breast tumor in ultrasound image, Biomedical Signal Processing and Control 8(6) (2013), 688–696.

Huang

, Tian

, Wang

, et al., Multi-objective flexible job-shop scheduling problem using modified discrete particle swarm optimization, Springer Plus 5(1) (2016), 1–22.

Zhang

, Xie

, Wu

, et al., Medical image classification using synergic deep learning, , Medical Image Analysis 54 (2019), 10–19.

Razzak

M.I.

, Naz

and Zaib

, Deep learning for medical image processing: Overview, challenges and the future, Classification in BioApps (2018), 323–350.

Shen

, Wu

and Suk

H.I.

, Deep learning in medical image analysis, , Annual Review of Biomedical Engineering 19 (2017), 221–248.

10.

, Yu

, Gu

, et al., Deep learning in genomic and medical image data analysis: challenges and approaches. (2), Journal of Information Processing Systems 13 (2017), 204–214.

11.

Bar

, Diamant

, Wolf

, et al., Chest pathology identification using deep feature selection with non-medical training, (3), Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization 6 (2018), 259–263.

12.

, Wang

, Guo

, et al., Developing and verifying automatic detection of active pulmonary tuberculosis from multi-slice spiral CT images based on deep learning, (5), Journal of X-ray Science and Technology 28 (2020), 939–951.

13.

Srivastava

and Purwar

R.K.

, Classification of CT scan images of lungs using deep convolutional neural network with external shape-based features, (1), Journal of Digital Imaging 33 (2020), 252–261.

14.

Gao

and Almekkawy

, ASU-Net++: A nested U-Net with adaptive feature extractions for liver tumor segmentation, Computers in Biology and Medicine 136 (2021), 104688.

15.

Polat

and Danaei Mehr

, Classification of pulmonary CT images by using hybrid 3D-deep convolutional neural network architecture, Applied Sciences 9(5) (2019), 940.

16.

Yap

M.H.

, Goyal

, Osman

, et al., Breast ultrasound region of interest detection and lesion localisation, , Artificial Intelligence in Medicine 107 (2020), 101880.

17.

Di Ruberto

, Loddo

and Putzu

, Detection of red and white blood cells from microscopic blood images using a region proposal approach, Computers in Biology and Medicine 116 (2020), 103530.

18.

Taghanaki

S.A.

, Abhishek

, Cohen

J.P.

, et al., Deep semantic segmentation of natural and medical images: a review, (1), Artificial Intelligence Review 54 (2021), 137–178.

19.

Wang

, Fan

and Wang

, Comparative analysis of image classification algorithms based on traditional machine learning and deep learning, Pattern Recognition Letters 141 (2021), 61–67.

20.

Jeyaraj

P.R.

and Nadar

E.R.S.

, Computer-assisted medical image classification for early diagnosis of oral cancer employing deep learning algorithm, Journal of Cancer Research and Clinical Oncology 145(4) (2019), 829–837.

21.

Melendez

, van Ginneken

, Maduskar

, et al., A novel multiple-instance learning-based approach to computer-aided detection of tuberculosis on chest x-rays, (1), IEEE Transactions on Medical Imaging 34 (2014), 179–192.

22.

, Wu

, Peng

, Liang

and Kong

, Automatic 3D liver segmentation based on deep learning and globally optimized surface evolution, Physics in Medicine & Biology 61(24) (2016), 8676.

23.

Esteva

, Kuprel

, Novoa

R.A.

, et al., Dermatologist-level classification of skin cancer with deep neural networks, , Nature 542 (2017), 115–118.

24.

Caicedo

J.C.

, Roth

, Goodman

, et al., Evaluation of deep learning strategies for nucleus segmentation in fluorescence images, (9), Cytometry Part A 95 (2019), 952–965.

25.

, Su

, Zhu

, et al., Neuron segmentation based on CNN with semi-supervised regularization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (2016), 20–28.

26.

Iqbal

, Ghani

M.U.

, Saba

, et al., Brain tumor segmentation in multi-spectral MRI using convolutional neural networks (CNN), Microscopy Research and Technique 81(4) (2018), 419–427.

27.

Durkee

M.S.

, Sibley

, Ai

, et al., Improved instance segmentation of immune cells in human lupus nephritis biopsies with Mask R-CNN, Proc SPIE (2020), 1132019.

28.

, Fang

and Lai

, Left ventricle automatic segmentation in cardiac MRI using a combined CNN and U-net approach, Computerized Medical Imaging and Graphics 82 (2020), 101719.

29.

Tang

, Zou

, Yang

, et al., A two-stage approach for automatic liver segmentation with Faster R-CNN and DeepLab, Neural Computing and Applications (2020), 1–10.

30.

Elshaer

M.E.A.

, Wisdom

and Mishra

, Transfer learning from sound representations for anger detection in speech. arXiv preprint arXiv:1902.02120 (2019).

31.

, Liu

and Wen Chen

, A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017), 4535–4544.

32.

Pak

and Kim

, A review of deep learning in image recognition. In 2017 4th International Conference on Computer Applications and Information Processing Technology (CAIPT) (2017, August), 1–3.

33.

George-Jones

N.A.

, Wang

, et al., Automated detection of Vestibular Schwannoma growth using a two-dimensional U-Net convolutional neural network, The Laryngoscope 131(2) (2021), E619–E624.

34.

Zhou

X.Y.

and Yang

G.Z.

, Normalization in training U-Net for 2-D biomedical semantic segmentation, IEEE Robotics and Automation Letters 4(2) (2019), 1792–1799.

35.

Baessler

, Luecke

, Lurz

, et al., Cardiac MRI and texture analysis of myocardial T1 and T2 maps in myocarditis with acute versus chronic symptoms of heart failure, (3), Radiology 292 (2019), 608–617.

36.

Ushinsky

, Bardis

, Glavis-Bloom

, et al., A 3D-2D hybrid U-net convolutional neural network approach to prostate organ segmentation of multiparametric MRI, (1), American Journal of Roentgenology 216 (2021), 111–116.

37.

Liu

, Song

Y.Q.

, Sheng

V.S.

, et al., Liver CT sequence segmentation based with improved U-Net and graph cut, , Expert Systems with Applications 126 (2019), 54–63.

38.

Kamnitsas

, Chen

, Ledig

, et al., Multi-scale 3D convolutional neural networks for lesion segmentation in brain MRI, Proc Ischemic Stroke Lesion Segmentation (2015), 13–16.

39.

Zhong

, Kim

, Plichta

, et al., Simultaneous cosegmentation of tumors in PET-CT images using deep fully convolutional networks, Medical Physics 46(2) (2019), 619–633.

40.

Xue

, Guo

and Ni

, Left ventricle quantification with sample-level confidence estimation via Bayesian neural network, Computerized Medical Imaging and Graphics 84 (2020), 101753.

41.

Zhou

, Siddiquee

M.M.R.

, Tajbakhsh

, et al., Unet++: A nested u-net architecture for medical image segmentation. In Deep learning in medical image analysis and multimodal learning for clinical decision support. Springer, Cham (2018), 3–11.

42.

Dias

P.A.

and Medeiros

, Semantic segmentation refinement by monte carlo region growing of high confidence detections. In Asian Conference on Computer Vision. Springer, Cham (2018), 131–146.

43.

Kumar

S.S.

, Moni

R.S.

and Rajeesh

, Automatic liver and lesion segmentation: a primary step in diagnosis of liver diseases, Signal, Image and Video Processing 7(1) (2013), 163–172.

44.

Budak

Ü.

, Guo

, Tanyildizi

, et al., Cascaded deep convolutional encoder-decoder neural networks for efficient liver tumor segmentation, Medical Hypotheses 134 (2020), 109431.

45.

Almotairi

, Kareem

, Aouf

, et al., Liver tumor segmentation in CT scans using modified SegNet, Sensors 20(5) (2020), 1516.

46.

Tian

, Liu

, Shi

, et al., Automatic Couinaud segmentation from CT volumes on liver using GLC-UNet. In International Workshop on Machine Learning in Medical Imaging. Springer, Cham (2019), 274–282.

47.

Qassim

, Verma

and Feinzimer

, Compressed residual-VGG16 CNN model for big data places image recognition. In 2018 IEEE 8th Annual Computing and Communication Workshop and Conference (CCWC) (2018), 169–175.

48.

Ananth

, Senthilkani

A.S.

, Gomathy

S.K.

, et al., Color image segmentation using IMOWT with 2D histogram grouping, (5), International Journal of Computer Science and Mobile Computing (IJCSMC) 3 (2014).

49.

Alom

M.Z.

, Yakopcic

, Hasan

, et al., Recurrent residual U-Net for medical image segmentation, Journal of Medical Imaging 6(1) (2019), 014006.

50.

, Wang

, Shi

, et al., Automatic liver segmentation in CT images based on support vector machine. In Proceedings of 2012 IEEE-EMBS International Conference on Biomedical and Health Informatics (2012), 333–336.

51.

Albishri

A.A

, Shah

S.J.H.

and Lee

, CU-Net: Cascaded U-Net model for automated liver and lesion segmentation and summarization. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (2019), 1416–1423.

U-Net: A valuable encoder-decoder architecture for liver tumors segmentation in CT images

Abstract

Keywords

1 Introduction

2.2 Data augmentation

2.3 Hepatic metastases tumor segmentation phase

2.3.1 Encoder–Decoder Architecture

Table 4 Summary of some studies reporting some liver lesion segmentation approaches. Literature Year Used techniques Used Dataset Dice (%) Lu et al. [50] 2012 SVM CT 85.38 Albishri et al. [51] 2019 U-net segmentation CT 89.64 Proposed method U-Net CT 93.54

References

Table 4
Summary of some studies reporting some liver lesion segmentation approaches.

Literature Year Used techniques Used Dataset Dice (%)

Lu et al. [50] 2012 SVM CT 85.38

Albishri et al. [51] 2019 U-net segmentation CT 89.64

Proposed method U-Net CT 93.54