![ ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091
3084
exhibited that the model achieved a promising dice for each case score of 69.79, indicating its effectiveness
in the breakdown process [1]. Additionally, the researchers found that SegNet outperformed U-Net in liver
segmentation, achieving a dice score of 95.53. Furthermore, the research demonstrated that incorporating an
additional classifier for tumor recognition significantly improved the results, particularly in the separation of
irregular tissues. This finding adds further validation to the efficacy of utilizing U-Net for lesion separation
tasks involving tumor-ridden livers [2]. The authors emphasized importance of automatic liver lesion
segmentation for achieving effective treatment outcomes and assisting medical experts. They proposed a
cascaded system utilizing both 3D and 2D “convolutional neural networks (CNN)” to segment hepatic
lesions. To appraise the segmentation results, a two-fold cross-validation was conducted on the liver tumor
segmentation benchmark (LiTS) dataset, aiming to identify any possible issues associated to
under-segmentation or over-segmentation [3].
This work presents a innovative technique for identification of liver cancer lumps and analyzing
their severity automatically. The usage of hybrid traits, rendering to the researchers, could help identify
malignant spots. The methodology encompasses some steps, opening with pre-processing analysis using
median filtering. This is followed by binary segmentation based on dynamic thresholding, and identification
of the “region of interest” (ROI) through the application of morphological functions. The simulation
outcomes demonstrate that the proposed model enhances the precision of detection with minimal
computational overhead. The achieved accuracy is reported as 92.67%, and the average detection time is 1.13
seconds. These results highlight the efficiency of the projected model in accurately detecting cancer tumors
[4]. In this research paper, the researchers propose an optimization technique designed for the automatic
recognition of tumor in abdominal liver images. These techniques significantly improve the efficacy of tumor
segmentation, contributing to more accurate detection. Notably, the water shed algorithm is specifically
employed in this analysis, yielding an impressive average accuracy of 0.97. Overall, this paper presents an
optimization technique that offers automated cancer detection in liver metaphors of abdomen [5]
Machine learning has emerged as a prevailing tool in biomedical imaging, particularly in the
medicinal tomography field. By employing machine learning methods for disease detection and cancer cell
identification, researchers have produced excellent results. In this study, the focus is on femur segmentation.
The researchers employed femur segmentation, a process that involves delineation and identification of the
femur bone in medical images. The projected technique yielded hopeful results, a dice value of 0.95 and with
a quick processing period of 0.93 seconds [6].
Experimental work offered here goals to adapt a deep learning prototye used for semantic
segmentation of road scenes to the segmentation of CT liver scan tumors in digital imaging and
communications in medicine (DICOM) format. With SegNet, image-level classification is accomplished by
pixel-level features using the trained VGG-16 image classification network as encoder and decoding
architecture as decoder. As opposed to conventional auto-encoders, SegNet saves only the max-pooling
indexes of feature maps instead of the entire maps. Most tumor parts can be correctly perceived by the
proposed method with an accuracy rate of over 86%. Although some false positives could be lowered by
applying false positive filters and training the model with more data, based on results, it appears that some
could be reduced by applying false positive filters [7]
The network architecture and training cases were optimized to generate a customCNN, and the final
network consisted of three convolutional layers with remedied linear units, two maximum pooling and 2 fully
linked layers. 494 hepatic lesions with typical imaging features were used in total, divided into training
(n=434) and testing (n=60). Cross-validation with Monte Carlo was used. Following the completion of model
engineering, the classification accurateness of the final CNN was compared with two board-certified
radiologists on an identical unseen test set. The DLS achieved 92% accuracy, 92% sensitivity (Sn), and 98%
specificity (Sp). In a single run of random unseen cases, test set performance averaged 90% Sn and 98% Sp.
For radiologists, the average Sn/Sp on these same cases was 82.5%/96.5%. The results showed a 90% Sn for
classifying hepatocellular carcinoma (HCC) compared to 60%/70% for radiologists. False and true positive
rates for HCC classification were 1.6% and 93.6%, respectively with a receiver working characteristic extent
under the curvature of 0.992 [8].
Using 494 lesions on multi-phasic magnetic resonance imaging (MRI), as discussed, a CNN was
created and trained to classify six hepatic tumour entities. Up to four important imaging features per lesion
were used to label a portion of each lesion class. Additionally, each detected characteristic received a
relevance score indicating its relative importance in the projected lesion classification. In identifying the
correct radiological features present in each test lesion, the interpretable deep learning system achieved
76.5% positive prognostic value and 82.9% sensitivity. The model misclassified 12% of lesions.
Misclassified lesions had more incorrect features than correctly identified lesions (60.4% vs. 85.6%). The
feature maps matched original image voxels that contributed to each imaging feature. Most significant
imaging standards for each class were reflected in the feature relevance scores [9].](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-2-320.jpg)
![Int J Artif Intell ISSN: 2252-8938
Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil)
3085
With increasing analysis applying CNNs in carcinoma pictures there's a requirement to value
rigorously their accuracy and outline future research directions. whereas current studies have lined major
liver cancers, the quantity of studies conducted up to now is little and limited, and a lot of research is required
to answer questions about the accurateness and compassion of CNN algorithms. However, many deficiencies
in current studies were observed. This be particularly vital in sight of the growing demand of CNNs in liver
medical specialty. Azer [10] propose a procedure for tumor breakdown of liver that involves several stages,
from preliminary to ultimate stage. The process begins with preprocessing steps, including grayscale
translation and median filtering, to eliminate clatter from the input images. Then, CNN is employed to sector
the liver section from CT images. Following the segmentation, the segmented CT volumes undergo feature
mining using procedures such as gray level co-occurrence matrix (GLCM) shape features, and “local binary
patterns (LBP)”. These features capture important appearances of the lump regions.
In this research analysis the research work introduces a novel approach for generating distinct
categories of polyp intrants, providing valuable evidence of liver sections in CT images. This method
incorporates a super pixel breakdown technique, enabling the mining of crucial and meaningful information
pertaining to the liver region. By reducing redundant information within the liver regions, the effectiveness of
the obtained information is enhanced, resulting in reduced computational complexity. In addition, the
abstraction of liver tumor cells was carried out with remarkable efficiency through the utilization of the
GLCM, shape features, and LBP. These advanced techniques, combined with application of active
contouring, facilitated the precise refinement of the extracted tumor cells.
3. METHOD
Figure 1 demonstrates the comprehensive flowchart representing segmentation of anticipated liver
tumour process. The process instigates with pre-processing of images, involving crucial steps such as
grayscale translation and median filtering. Subsequently, part of the liver is segmented using a CNN,
ensuring accurate delineation.
Figure 1. Flow chart: Feasible technique
Step 1: CT scans undergo preprocessing,which involves procedures such as grayscale translation and median
filtering.
Recent automated deep learning-based techniques, designed without incorporating pre-processing
analysis or similar approaches, may not exhibit the desired level of reliability when applied to larger
databases or extensive timeframes. It is well-established that CT volumes contain various types of noise, such
as impulse, and gaussian quantization noise. The initial step focuses on fine-tuning low-contrast CT images
through an intensity value tuning function, which aims to mitigate noise-related challenges and improve
complete recital of deep learning model.](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-3-320.jpg)
![ ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091
3086
X = imadjust (Y) (1)
X = imadjust (Y, [L_in, H_in]) (2)
X = imadjust (Y, [L_in, H_in], [L_out, H_out]) (3)
In (1) enhances the dissimilarity of output image X by mapping intensity values from grayscale image. Y to
new values within the range of b. By evasion, the imadjust function saturates the top 1% and bottom1% of
all pixel values. In (2), L_in and H_in denote the lower and higher input values, respectively, both mapped to
0 and 1. In (3) extends this idea by mapping values between L_in and H_in to the range between L_out and
H_out, aiming to minimize noise and artifacts effectively.
Step 2: after eliminating noise in the previous step, we employ a CNN to segment the CT volumes and
extract the liver region [11].
Our method employs CNN to implement time implied phase growth for liver recognition and
segmentation, representing a notable advancement over prior techniques using CNN for this purpose.
Figure 2 demonstrates the utilization of convolutional filters and neural network connections in our approach,
crucial for precise and efficient liver analysis. “A CNN method tailored for 2D abdominal liver segmentation
was proposed to ensure precision in segmenting liver from CT scan images. This approach integrates
classifiers and creators to delineate "non-liver" and "liver" regions using two-way layers with softmax
probabilities. The CNN architecture comprises five layers, amalgamating information from five input images,
while subsequent layers employ fully connected neural networks or max-pooling techniques to refine
features. Employing canonical residual networks (ResNet) as encoders, global convolutional networks
(GCNs) for decoding, and PatchGANs as classifiers, the system achieves high-resolution segmentation.
DICOM images, acquired noncontinuously in an axial manner, are converted to neuroimaging informatics
technology initiative (NIFTI) format to reconstruct 3D volumes. Additionally, each axial slice within a 3D
CT volume is resized to 256 by 256 for testing and training purposes. The subsequent section elaborates on
identifying malignant tumors through tumor candidate generation in segmented regions, with Figure 3
displaying obtained segmentation outcomes.
Figure 2. Convolution neural network architecture
Figure 3. Segmentation results (green outline indicates physically recognized boundary and red outline
signifies automatically recognized boundary)
Step 3: to efficiently extract the ROI from the segmented liver part, we utilize the GLCM-CNN approach.
This method focuses on capturing the distribution of grey levels, which represents the illumination
values of pixels within an image. The GLCM technique is employed to gather information about image
texture structures. Specifically, the GLCM considers spatial associations between two pixels at a specific
orientation angle and distance. The GLCM matrix function is obtained by calculating the occurrence of pixel
pairs with a certain distance (D) and orientation angle (θ), such as 0, 45, 90, and 120 degrees. This work
utilizes several properties of the GLCM, namely homogeneity, energy, and contrast [12]. Homogeneity, also
known as the inverse difference moment, is the reciprocal of contrast and measures the distribution of](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-4-320.jpg)
![Int J Artif Intell ISSN: 2252-8938
Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil)
3087
elements along the GLCM diagonal. In (4) can be used to calculate homogeneity. Energy, also referred to as
uniformity, quantifies the concentration of gray levels in the GLCM. It represents the number of squared
elements in the GLCM and can be computed using (5) and (6) can be used to calculate contrast. In (4)-(6),
the vertical and horizontal coordinates in the GLCM matrix are represented by 'a' and 'b', respectively, and
the matrix value at coordinate (a, b) is denoted by M (a, b).
𝐻 =
∑ 𝑀𝑎,𝑏
𝑋
𝑎,𝑏=1
1+(𝑎+𝑏)2
(4)
𝐸 = ∑ 𝑀𝑎 ,𝑏2
𝑋
𝑎 ,𝑏=1 (5)
𝐶 = ∑ 𝑀𝑎,𝑏 (𝑎 − 𝑏)2
𝑋
𝑎 ,𝑏=1 (6)
By incorporating GLCM texture features into the input layer, an improved technique of the CNN is
presented, as depicted in Figure 4, which showcases the architecture of GLCM-CNN. In the first layer, an
image of size 228×344 pixels are obtained from features extracted using the GLCM method. It is calculated
based on the major axis length (a) and minor axis length (b) and is represented by (7). Perimeter is another
feature that determines the boundary span of the ROI and generates an array where pixels with 255 values
indicate the “border and pixels with 0 values indicate the interior. In (8) represents the perimeter calculation,
with A and B representing the edges of the ROI. Additionally, trajectory represents the arrangement of pixels
(ath and bth) by generating corresponding curvatures. Lastly, the important feature of area is calculated using
(9), where a and b represent the number of pixels within the shape. An ROI vector is constructed, consisting
of a ROI at position A and a ROI at position B [13], [14]. These GLCM texture features contribute to the
overall classification performance of the GLCM-CNN model.
𝐸 =
𝑎
𝑏
(7)
𝑃 = (𝑃𝑎 ,𝑏, 𝐴 𝑒𝑑𝑔𝑒[𝑃] = 𝑎, 𝐵 𝑒𝑑𝑔𝑒 [𝑃] = 𝑏) (8)
𝐴 = (𝐴𝑎,𝑏 , 𝐴 𝑅𝑂𝐼[𝐴𝑟𝑒𝑎] = 𝑎, 𝐵 𝑅𝑂𝐼 [𝐴𝑟𝑒𝑎] = 𝑏) (9)
The modified technique depicted in Figure 4 is referred to as GLCM-CNN, with the architecture
consisting of sub-sampling layers and three types of layers: convolutional layers followed by an output layer.
The purpose of using the GLCM is to extract features that characterize a set of structures, this helps decrease
the misclassification of cancerous tumors in images. In (10) outlines the sub-sampling function, wherein the
input patch window function denotes the neighborhood X(n, n), and calculates the average value of this
neighborhood. The output layer is directly linked to the last convolutional layer and comprises four to five
neurons, each representing various stages of cancer. This layer produces the ultimate output utilizing th e
acquired features, thereby facilitating the classification process.
𝐴𝑗 = (𝐴𝑗
𝑛∗𝑛
∗ 𝑋(𝑛, 𝑛))
𝑛∗𝑛
𝑎𝑣𝑔
(10)
Figure 4. Architecture of GLCM-CNN
During CNN training, the objective is to minimize the average squared error obtained from the training data.
In (11) defines the average squared error, where X represents the total number of training data samples,
d_j(n) denotes the target of the actual class, B stands for the number of batches, y_j(n) signifies the output of
the jth layer, and E_average indicates the normalized value of the squared error.
Eaverage =
1
X
∑ (dj
(n) − yj
(n))2
B
a=1 (11)](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-5-320.jpg)
![ ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091
3088
Step 4: Active contour method is utilized for refining the identification of liver tumors [15]
Anil and Dayananda [11] presented a comprehensive convex active contour model (ACM) for reflex
image segmentation to refine the boundaries obtained in step 3 of our methodology. This technique involves
three main steps: contour Initialization, regional term formulation, and boundary term formulation. The ACM
is utilized to enhance and fine-tune the tumor boundaries.It aims to recover lost borders or eliminate unwanted
borders from the image. The function hr is a region function that evaluates the exterior and interior regions
based on specific constraints.The function u operates on the image domain and assigns a value between 0 and
1 to each pixel position x in the image. For a detailed explanation of this step, please refer to [11].
𝑚𝑖𝑛0 ≤𝑢≤1(∫ 𝑔𝑏
|∇𝑢|𝑑𝑥 + 𝜆 ∫ ℎ𝑟𝑢𝑑𝑥
𝛺
𝛺 ) (12)
4. RESULTS AND DISCUSSION
In this section, we conducted a comprehensive comparison and evaluation of various segmentation
techniques vis-à-vis our proposed method. Through meticulous analysis and comparison with other
methodologies [16]‒[19], we have established that our segmentation approach outperforms these methods,
exhibiting a notable improvement in dice precision ranging from 6% to 10%. In (13) delineates the accuracy
calculation. It's important to note that the specific equations and intricacies regarding accuracy computation
may vary depending on the methodology elucidated in the original papers [16]‒[19]. For a more thorough
understanding, we recommend referring to the original research papers for a detailed explanation of the
accuracy calculation methodology.
A =
No.of correct identification
total number of test data
𝑋 100 (13)
Table 2 also features the maximum and minimum values for each category of tumors. It's crucial to
emphasize that without access to the actual table and specific data, we unable to provide exact scores and
comparisons for each metric. Figures 5 and 6 provides a quantitative evaluation of the proposed segmentation
technique, comparing it to CNN and GLCM-CNN methods. Analysis from the table reveals that the
GLCM-CNN method achieves higher accuracy. Table 3 shows the classification precision of two models: CNN
and GLCM-CNN. Table 4 further presents an assessment of the proposed segmentation procedure, analyzing
segmentation results for small, medium, and large tumors. Metrics such as relative volume differences (RVD),
dice per case, volume overlap error (VOE), average symmetric difference (ASD), and maximum symmetric
difference (MSD) are utilized to quantify the scores [20]‒[22]. Nevertheless, this table facilitates a thorough
assessment of the segmentation results, offering valuable insights into the performance of the proposed
segmentation technique [23]‒[25] across various tumor sizes. For detailed information and precise values.
Table 2. Segmentation results comparison with other approaches
Method/Author Dice per case ASD (mm) RVD VOE MSD (mm)
Chris et al. [16] 0.56 - - - -
Yuan et al. [17] 0.56 1.151 0.228 0.378 6.269
Chlebus et al. [18] 0.65 11.11 0.380 0.625 16.71
Chen et al. [6] 0.67 6.36 0.41 0.564 11.69
Proposedmethod[CNN+GLCM] 0.86 4.753 0.234 0.312 10.21
Figure 5. Graphical representation of classification
precision of CNN and GLCM-CNN
Figure 6. Graphical illustration of scores (Table 4)](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-6-320.jpg)
![Int J Artif Intell ISSN: 2252-8938
Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil)
3089
Table 3. Classification precision of CNN and GLCM-CNN
Iterations CNN GLCM-CNN
H E C
25 66.78 52.16 79.65 88.69
50 77.25 55.24 86.63 88.62
100 86.34 69.23 93.65 89.72
200 89.89 89.91 94.65 91.76
Table 4. Evaluation of the proposed tumor segmentation technique quantitatively
DICE VOE (%) RVD (%) ASD (mm) MSD (mm)
Small tumor 0.87 ± 2.25 36.91 ± 8.26 24.69 ± 5.36 0.76 ± 0.21 0.98 ± 0.30
Max 0.87 44.65 29.31 0.91 0.92
Min 0.71 24.12 20.62 0.54 0.69
Medium tumor 0.81 ± 0.25 37.62 ± 9.62. 22.54 ± 3.72 0.73 ± 0.51 1.14 ± 0.52
Max 0.91 46.03 25.56 0.85 0.86
Min 0.73 32.45 19.62 0.56 0.68
Large tumor 0.91 ± 1.25 38.95 ± 3.62 19.23 ± 2.56 0.78 ± 0.32 1.12 ± 0.84
Max 0.94 41.53 21.79 0.92 0.85
Min 0.75 35.31 17.23 0.52 0.67
Average 0.87 ± 7.06 28.05 ± 4.32 23.25 ± 2.95 0.77 ± 0.56 1.11 ± 0.21
5. CONCLUSION
Liver cancer, the deadliest form globally, demands precise analysis and swift tumor identification
across all stages to improve survival rates. A new method, using CT scan images, is being developed to
detect liver cancer stages efficiently. Initial steps involve image preprocessing, converting to grayscale and
applying median filtering. A novel segmentation technique, including CNN, accurately segments cancerous
polyps. Key properties like eccentricity, perimeter, and area are extracted using the GLCM method from the
segmented ROI. The ACM refines tumor localization. This combined approach has shown promising results,
achieving a dice similarity coefficient of 81.8 ± 14.5% during training and 80.6 ± 12.9% during testing on CT
images from 132 patients.
APPENDIX
Table 1. Different stages of cancer
Stage Stage grouping Stage description
1 C0 C0 indicates No evidence of a primary tumor.
1A C1a
L0
D0
C1a indicates not grown to blood vessels and it is only tumor size varies from 1-2 cm.
L0 indicates has not spread to nearby lymph nodes.
D0 indicates has not spread to distant sites.
1B C1b
L0
D0
C1b indicates not grown to blood vessels and it is only tumor size varies from 2-4 cm.
L0 indicates has not spread to nearby lymph nodes.
D0 indicates has not spread to distant sites.
2 C2
L0
D0
C2 indicates may be growth of single tumor, size is great than 2 cm and parallelly has grown into
blood vessels or may be grown of more than one tumor but not larger than 6cm.
L0 indicates has not spread to nearby lymph nodes.
D0 indicates has not spread to distant sites.
3A C3
L0
D0
C3 indicates growth of more than one tumor, size is great than 6 cm
L0 indicates has not spread to nearby lymph nodes.
D0 indicates has not spread to distant sites.
3B C4
L0
D0
C4 indicates one tumorof somewhat size that has grown-up intoa mainbranchof a large vein of the.
L0 indicates has not spread to nearby lymph nodes.
D0 indicates has not spread to distant sites.
4A Any C
L1
D0
L1 indicates growthof single tumour ormultipletumors ofanysize has blowout toneighboringlymph
nodes.
D0 indicates has not blowout to detached sites.
4B Any C
Any L
D1
D1 indicates It has spread to detached organs such as the bones or lungs.
REFERENCES
[1] E. Smith, “Worldcancerday 2017: livercancer, a global challenge thanks toviruses and alcohol,” Cancer Research UK, 2017,
Accessed: Jun. 10, 2018. [Online]. Available: https://news.cancerresearchuk.org/2017/02/13/world-cancer-day-2017-liver-cancer-
a-global-challenge-thanks-to-viruses-and-alcohol/
[2] N. Nanda, P. Kakkar, andS. Nagpal, “Computer-aidedsegmentationof liver lesions in CT scans using cascaded convolutional](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-7-320.jpg)
![ ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091
3090
neural networks andgeneticallyoptimisedclassifier,”ArabianJournal for Scienceand Engineering, vol. 44, no. 4, pp. 4049–
4062, 2019, doi: 10.1007/s13369-019-03735-8.
[3] R. Dey and Y. Hong, “Hybrid cascaded neural network for liver lesion segmentation,” in 2020 IEEE 17th International
Symposium on Biomedical Imaging (ISBI), 2020, pp. 1173–1177, doi: 10.1109/ISBI45749.2020.9098656.
[4] S. I. TurabandV. K. Kadam, “Livercancer detection and grading using efficient computer vision techniques,” Annals of the
Romanian Society for Cell Biology, vol. 25, no. 2, pp. 1740–1755, 2021.
[5] A. Krishan andD. Mittal,“Ensembledlivercancer detection andclassification usingCT images,” Proceedings of the Institution
of Mechanical Engineers, Part H: Journal of Engineering in Medicine, vol. 235, no. 2, pp. 232-244, 2021, doi:
10.1177/0954411920971888.
[6] F. Chen, J. Liu, Z. Zhao, M. Zhu, andH. Liao, “Three-dimensional feature-enhancednetworkforautomatic femur segmentation,”
IEEE Journal of Biomedical andHealth Informatics, vol. 23, no. 1, pp. 243–252, 2019, doi: 10.1109/JBHI.2017.2785389.
[7] S. Almotairi,G. Kareem, M.Aouf, B. Almutairi, andM. A. M. Salem, “Liver tumor segmentation in CT scans using modified
segnet,” Sensors, vol. 20, no. 5, 2020, doi: 10.3390/s20051516.
[8] C. A. Hamm et al., “Deeplearningfor livertumordiagnosis part I: development of a convolutional neural network classifier for
multi-phasic MRI,” European Radiology, vol. 29, no. 7, pp. 3338–3347, 2019.
[9] C. J. Wanget al., “Deeplearningfor liver tumordiagnosis part II: convolutionalneural network interpretation using radiologic
imaging features,” European Radiology, vol. 29, no. 7, pp. 3348–3357, 2019, doi: 10.1007/s00330-019-06214-8.
[10] S. A. Azer, “Deep learningwith convolutional neural networks foridentificationof liver masses andhepatocellular carcinoma: A
systematic review,” World Journal of Gastrointestinal Oncology, vol. 11, no. 12, pp. 1218–1230, 2019, doi:
10.4251/wjgo.v11.i12.1218.
[11] B. C. Anil andP. Dayananda, “Automatic livertumorsegmentation basedon multi-level deep convolutional networks and fractal
residual network,” IETEJournal of Research, vol. 69, no. 4, pp. 1925–1933, 2023, doi: 10.1080/03772063.2021.1878066.
[12] U. Hameed, M. U. Rehman, A. Rehman, R. Damaševičius, A. Sattar, and T. Saba, “A deep learning approach for liver cancer
detectionin CT scans,” Computer Methods in Biomechanics and Biomedical Engineering: Imaging& Visualization, vol. 11, no.
7, 2024, doi: 10.1080/21681163.2023.2280558.
[13] X. Jia andM. Q. H. Meng, “A deep convolutional neural networkforbleedingdetection in wireless capsule endoscopy images,”
in 2016 38thAnnual International Conferenceof the IEEE Engineeringin Medicine andBiology Society (EMBC), 2016, pp. 639–
642, doi: 10.1109/EMBC.2016.7590783.
[14] K. Suzuki, J. Shiraishi, H. Abe, H. MacMahon, K. Doi, “False-positive reduction in computer-aided diagnostic scheme for
detectingnodules in chest radiographs by means ofmassivetrainingartificial neuralnetwork,” AcademicRadiology, vol. 12, no.
2, pp. 191–201, 2005, doi: 10.1016/j.acra.2004.11.017.
[15] M. Kass, A. Witkin,andD. Terzopoulos, “Snakes: active contour models,” International Journal of Computer Vision, vol. 1, no.
4, pp. 321–331, 1988, doi: 10.1007/BF00133570.
[16] P. F. Christ et al., “Automatic liver andtumorsegmentation ofCT andMRI volumes using cascaded fully convolutional neural
networks,” Arxiv-Computer Science, vol. 1, pp. 1–20, 2017.
[17] Y. Yuan, “Hierarchical convolutional-deconvolutional neural networks for automatic liver and tumor segmentation,” Arxiv-
Computer Science, pp. 1–4, 2017.
[18] G. Chlebus, H. Meine, J. H. Moltz, andA. Schenk, “Neural network based automatic liver tumor segmentation with random
forest-based candidate ltering,” Arxiv-Computer Science, vol. 1, no. 1–4, 2017.
[19] X. Han, “Automatic liverlesionsegmentationusinga deep convolutional neural networkmethod,” Arxiv-Computer Science, vol.
1, pp. 1–4, 2017.
[20] K. He, X. Zhang, S. Ren, andJ. Sun, “Deep residual learningforimage recognition,” in 2016 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90.
[21] C. Peng, X. Zhang, G. Yu, G. Luo, andJ. Sun, “Large kernel matters--improve semantic segmentation by global convolutional
network,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4353–4361.
[22] P. Isola, J.-Y. Zhu, T.Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1125–1134.
[23] M. M. Santoni, D. I. Sensuse, A. M. Arymurthy,and M. I. Fanany, “Cattle race classification using gray level co -occurrence
matrix convolutional neural networks,” Procedia Computer Science, vol. 59, pp. 493–502, 2015, doi:
10.1016/j.procs.2015.07.525.
[24] S. S. Roy, S. Roy, P. Mukherjee,andA. H. Roy, “An automated liver tumour segmentation and classification model by deep
learningbasedapproaches,” Computer Methods in Biomechanics andBiomedical Engineering:Imaging& Visualization, vol. 11,
no. 3, pp. 638-650, 2023, doi: 10.1080/21681163.2022.2099300.
[25] M. Hussain,N. Saher, andS. Qadri, “Computer vision approachfor liver tumor classificationusingCT dataset,” Applied Artificial
Intelligence, vol. 36, no. 1, 2022, doi: 10.1080/08839514.2022.2055395.
BIOGRAPHIES OF AUTHORS
Bellary Chiterki Anil is working as Associate Professor & Head, Department of
Computer Science and Engineering (AI & ML), JSS Academy of Technical Education,
Bengaluru. He has completed B.E. degree in CSE Engineering (VTU) from RYMEC, Bellary,
Karnataka, M.Tech. in CSE from SBMJCE, Karnataka, and Ph.D. in CSE Engineering from
VTU, Belagavi in 2021. He has a teaching experience of more than 10 years. His area of
researchworks are data mining and image processing. He has published over 25 papers in
international journals and conferences. He is a member of CSI, ISTE societies in India. He is a
reviewer for International Journals and conferences. He can be contacted at email:
anil.bc2@gmail.com.](https://image.slidesharecdn.com/6225119-250522061655-ea4a63c0/85/Identifying-liver-cancer-cells-using-cascaded-convolutional-neural-network-and-gray-level-co-occurrence-matrix-techniques-8-320.jpg)
Identifying liver cancer cells using cascaded convolutional neural network and gray level co-occurrence matrix techniques
- 1. IAES International Journal of Artificial Intelligence (IJ-AI) Vol. 13, No. 3, September 2024, pp. 3083~3091 ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i3.pp3083-3091 3083 Journal homepage: http://ijai.iaescore.com Identifying liver cancer cells using cascaded convolutional neural network and gray level co-occurrence matrix techniques Bellary Chiterki Anil1, Arun Kumar Gowdru2, Dayananda Prithviraja3, Niranjan Chanabasappa Kundur4, Balakrishnan Ramadoss5 1 Department of Computer Science and Engineering (AI & ML), JSS Academy of Technical Education, Bengaluru, India 2 Department of Electronics and Communication Engineering, JSS Academy of Technical Education, Noida, India 3 Department ofInformationTechnology,Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India 4 Department of Computer Science and Engineering, JSS Academy of Technical Education, Bengaluru, India 5 Department of Computer Applications, National Institute of Technology, Tiruchirapalli, India Article Info ABSTRACT Article history: Received Jan 30, 2024 Revised Feb 20, 2024 Accepted Feb 28, 2024 Liver cancer has a high mortality rate, especially in South Asia, East Asia, and Sub-Saharan Africa. Efforts to reduce these rates focus on detecting liver cancer at all stages. Early detection allows more treatment options, though symptoms may not always be apparent. The staging process evaluates tumor size, location, lymph node involvement, and spread to other organs. Our research used the CLD staging system, assessing tumor size (C), lymph nodes (L), and distant invasion (D). We applied a deep learning approach with a cascaded convolutional neural network (CNN) and gray level co-occurrence matrix (GLCM)-based texture features to distinguish benign from malignant tumors. The method validated with the cancer imaging archive (TCIA) dataset, demonstrating superior accuracy compared to existing techniques. Keywords: Computed tomography Hepatocellular carcinoma Metastatic carcinoma Convolutional neural network Region of interest Machine learning This is an open access article under the CC BY-SA license. Corresponding Author: Arun Kumar Gowdru Department of Electronics and Communication Engineering, JSS Academy of Technical Education Noida, India Email: [email protected] 1. INTRODUCTION Liver tissue cancer, a malignancy characterized by the proliferation of cancerous cells within the liver, presents a significant challenge. The aimof this analysis is to accurately identify cancerous regions and estimate size of malignant tissue from “computed tomography (CT)” scan slices. These slices cover the liver and adjacent internal organs, spanning from top to bottom. Through segmentation of each slice, volumetric measurements of the entire liver can be obtained, facilitating the assessment of affected tissue extent at different stages of cancer progression, as detailed in Table 1 (see in Appendix). Liver cancer constitutes a health concern globally, with a high risk of recurrent occurrences. In 2020 alone, there were approximately 9.5 lakh newly diagnosed cases of liver cancer annually, ranking it as the fourth leading cause of cancer-related deaths worldwide across all income brackets. According to the WHO global cancer data, projections for 2019 indicated around 8.7 million new cases of cancer diagnosed globally, resulting in around 9.8 million deaths.", with liver cancer contributing to approximately 8.2%, or roughly 782,000 deaths. 2. RELATED WORK The researchers explored various convolutional network architectures for subdivision and tumor recognition purposes. The primary focus of the study was to assess the recital of U-Net and SegNet. Results
- 2. ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091 3084 exhibited that the model achieved a promising dice for each case score of 69.79, indicating its effectiveness in the breakdown process [1]. Additionally, the researchers found that SegNet outperformed U-Net in liver segmentation, achieving a dice score of 95.53. Furthermore, the research demonstrated that incorporating an additional classifier for tumor recognition significantly improved the results, particularly in the separation of irregular tissues. This finding adds further validation to the efficacy of utilizing U-Net for lesion separation tasks involving tumor-ridden livers [2]. The authors emphasized importance of automatic liver lesion segmentation for achieving effective treatment outcomes and assisting medical experts. They proposed a cascaded system utilizing both 3D and 2D “convolutional neural networks (CNN)” to segment hepatic lesions. To appraise the segmentation results, a two-fold cross-validation was conducted on the liver tumor segmentation benchmark (LiTS) dataset, aiming to identify any possible issues associated to under-segmentation or over-segmentation [3]. This work presents a innovative technique for identification of liver cancer lumps and analyzing their severity automatically. The usage of hybrid traits, rendering to the researchers, could help identify malignant spots. The methodology encompasses some steps, opening with pre-processing analysis using median filtering. This is followed by binary segmentation based on dynamic thresholding, and identification of the “region of interest” (ROI) through the application of morphological functions. The simulation outcomes demonstrate that the proposed model enhances the precision of detection with minimal computational overhead. The achieved accuracy is reported as 92.67%, and the average detection time is 1.13 seconds. These results highlight the efficiency of the projected model in accurately detecting cancer tumors [4]. In this research paper, the researchers propose an optimization technique designed for the automatic recognition of tumor in abdominal liver images. These techniques significantly improve the efficacy of tumor segmentation, contributing to more accurate detection. Notably, the water shed algorithm is specifically employed in this analysis, yielding an impressive average accuracy of 0.97. Overall, this paper presents an optimization technique that offers automated cancer detection in liver metaphors of abdomen [5] Machine learning has emerged as a prevailing tool in biomedical imaging, particularly in the medicinal tomography field. By employing machine learning methods for disease detection and cancer cell identification, researchers have produced excellent results. In this study, the focus is on femur segmentation. The researchers employed femur segmentation, a process that involves delineation and identification of the femur bone in medical images. The projected technique yielded hopeful results, a dice value of 0.95 and with a quick processing period of 0.93 seconds [6]. Experimental work offered here goals to adapt a deep learning prototye used for semantic segmentation of road scenes to the segmentation of CT liver scan tumors in digital imaging and communications in medicine (DICOM) format. With SegNet, image-level classification is accomplished by pixel-level features using the trained VGG-16 image classification network as encoder and decoding architecture as decoder. As opposed to conventional auto-encoders, SegNet saves only the max-pooling indexes of feature maps instead of the entire maps. Most tumor parts can be correctly perceived by the proposed method with an accuracy rate of over 86%. Although some false positives could be lowered by applying false positive filters and training the model with more data, based on results, it appears that some could be reduced by applying false positive filters [7] The network architecture and training cases were optimized to generate a customCNN, and the final network consisted of three convolutional layers with remedied linear units, two maximum pooling and 2 fully linked layers. 494 hepatic lesions with typical imaging features were used in total, divided into training (n=434) and testing (n=60). Cross-validation with Monte Carlo was used. Following the completion of model engineering, the classification accurateness of the final CNN was compared with two board-certified radiologists on an identical unseen test set. The DLS achieved 92% accuracy, 92% sensitivity (Sn), and 98% specificity (Sp). In a single run of random unseen cases, test set performance averaged 90% Sn and 98% Sp. For radiologists, the average Sn/Sp on these same cases was 82.5%/96.5%. The results showed a 90% Sn for classifying hepatocellular carcinoma (HCC) compared to 60%/70% for radiologists. False and true positive rates for HCC classification were 1.6% and 93.6%, respectively with a receiver working characteristic extent under the curvature of 0.992 [8]. Using 494 lesions on multi-phasic magnetic resonance imaging (MRI), as discussed, a CNN was created and trained to classify six hepatic tumour entities. Up to four important imaging features per lesion were used to label a portion of each lesion class. Additionally, each detected characteristic received a relevance score indicating its relative importance in the projected lesion classification. In identifying the correct radiological features present in each test lesion, the interpretable deep learning system achieved 76.5% positive prognostic value and 82.9% sensitivity. The model misclassified 12% of lesions. Misclassified lesions had more incorrect features than correctly identified lesions (60.4% vs. 85.6%). The feature maps matched original image voxels that contributed to each imaging feature. Most significant imaging standards for each class were reflected in the feature relevance scores [9].
- 3. Int J Artif Intell ISSN: 2252-8938 Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil) 3085 With increasing analysis applying CNNs in carcinoma pictures there's a requirement to value rigorously their accuracy and outline future research directions. whereas current studies have lined major liver cancers, the quantity of studies conducted up to now is little and limited, and a lot of research is required to answer questions about the accurateness and compassion of CNN algorithms. However, many deficiencies in current studies were observed. This be particularly vital in sight of the growing demand of CNNs in liver medical specialty. Azer [10] propose a procedure for tumor breakdown of liver that involves several stages, from preliminary to ultimate stage. The process begins with preprocessing steps, including grayscale translation and median filtering, to eliminate clatter from the input images. Then, CNN is employed to sector the liver section from CT images. Following the segmentation, the segmented CT volumes undergo feature mining using procedures such as gray level co-occurrence matrix (GLCM) shape features, and “local binary patterns (LBP)”. These features capture important appearances of the lump regions. In this research analysis the research work introduces a novel approach for generating distinct categories of polyp intrants, providing valuable evidence of liver sections in CT images. This method incorporates a super pixel breakdown technique, enabling the mining of crucial and meaningful information pertaining to the liver region. By reducing redundant information within the liver regions, the effectiveness of the obtained information is enhanced, resulting in reduced computational complexity. In addition, the abstraction of liver tumor cells was carried out with remarkable efficiency through the utilization of the GLCM, shape features, and LBP. These advanced techniques, combined with application of active contouring, facilitated the precise refinement of the extracted tumor cells. 3. METHOD Figure 1 demonstrates the comprehensive flowchart representing segmentation of anticipated liver tumour process. The process instigates with pre-processing of images, involving crucial steps such as grayscale translation and median filtering. Subsequently, part of the liver is segmented using a CNN, ensuring accurate delineation. Figure 1. Flow chart: Feasible technique Step 1: CT scans undergo preprocessing,which involves procedures such as grayscale translation and median filtering. Recent automated deep learning-based techniques, designed without incorporating pre-processing analysis or similar approaches, may not exhibit the desired level of reliability when applied to larger databases or extensive timeframes. It is well-established that CT volumes contain various types of noise, such as impulse, and gaussian quantization noise. The initial step focuses on fine-tuning low-contrast CT images through an intensity value tuning function, which aims to mitigate noise-related challenges and improve complete recital of deep learning model.
- 4. ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091 3086 X = imadjust (Y) (1) X = imadjust (Y, [L_in, H_in]) (2) X = imadjust (Y, [L_in, H_in], [L_out, H_out]) (3) In (1) enhances the dissimilarity of output image X by mapping intensity values from grayscale image. Y to new values within the range of b. By evasion, the imadjust function saturates the top 1% and bottom1% of all pixel values. In (2), L_in and H_in denote the lower and higher input values, respectively, both mapped to 0 and 1. In (3) extends this idea by mapping values between L_in and H_in to the range between L_out and H_out, aiming to minimize noise and artifacts effectively. Step 2: after eliminating noise in the previous step, we employ a CNN to segment the CT volumes and extract the liver region [11]. Our method employs CNN to implement time implied phase growth for liver recognition and segmentation, representing a notable advancement over prior techniques using CNN for this purpose. Figure 2 demonstrates the utilization of convolutional filters and neural network connections in our approach, crucial for precise and efficient liver analysis. “A CNN method tailored for 2D abdominal liver segmentation was proposed to ensure precision in segmenting liver from CT scan images. This approach integrates classifiers and creators to delineate "non-liver" and "liver" regions using two-way layers with softmax probabilities. The CNN architecture comprises five layers, amalgamating information from five input images, while subsequent layers employ fully connected neural networks or max-pooling techniques to refine features. Employing canonical residual networks (ResNet) as encoders, global convolutional networks (GCNs) for decoding, and PatchGANs as classifiers, the system achieves high-resolution segmentation. DICOM images, acquired noncontinuously in an axial manner, are converted to neuroimaging informatics technology initiative (NIFTI) format to reconstruct 3D volumes. Additionally, each axial slice within a 3D CT volume is resized to 256 by 256 for testing and training purposes. The subsequent section elaborates on identifying malignant tumors through tumor candidate generation in segmented regions, with Figure 3 displaying obtained segmentation outcomes. Figure 2. Convolution neural network architecture Figure 3. Segmentation results (green outline indicates physically recognized boundary and red outline signifies automatically recognized boundary) Step 3: to efficiently extract the ROI from the segmented liver part, we utilize the GLCM-CNN approach. This method focuses on capturing the distribution of grey levels, which represents the illumination values of pixels within an image. The GLCM technique is employed to gather information about image texture structures. Specifically, the GLCM considers spatial associations between two pixels at a specific orientation angle and distance. The GLCM matrix function is obtained by calculating the occurrence of pixel pairs with a certain distance (D) and orientation angle (θ), such as 0, 45, 90, and 120 degrees. This work utilizes several properties of the GLCM, namely homogeneity, energy, and contrast [12]. Homogeneity, also known as the inverse difference moment, is the reciprocal of contrast and measures the distribution of
- 5. Int J Artif Intell ISSN: 2252-8938 Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil) 3087 elements along the GLCM diagonal. In (4) can be used to calculate homogeneity. Energy, also referred to as uniformity, quantifies the concentration of gray levels in the GLCM. It represents the number of squared elements in the GLCM and can be computed using (5) and (6) can be used to calculate contrast. In (4)-(6), the vertical and horizontal coordinates in the GLCM matrix are represented by 'a' and 'b', respectively, and the matrix value at coordinate (a, b) is denoted by M (a, b). 𝐻 = ∑ 𝑀𝑎,𝑏 𝑋 𝑎,𝑏=1 1+(𝑎+𝑏)2 (4) 𝐸 = ∑ 𝑀𝑎 ,𝑏2 𝑋 𝑎 ,𝑏=1 (5) 𝐶 = ∑ 𝑀𝑎,𝑏 (𝑎 − 𝑏)2 𝑋 𝑎 ,𝑏=1 (6) By incorporating GLCM texture features into the input layer, an improved technique of the CNN is presented, as depicted in Figure 4, which showcases the architecture of GLCM-CNN. In the first layer, an image of size 228×344 pixels are obtained from features extracted using the GLCM method. It is calculated based on the major axis length (a) and minor axis length (b) and is represented by (7). Perimeter is another feature that determines the boundary span of the ROI and generates an array where pixels with 255 values indicate the “border and pixels with 0 values indicate the interior. In (8) represents the perimeter calculation, with A and B representing the edges of the ROI. Additionally, trajectory represents the arrangement of pixels (ath and bth) by generating corresponding curvatures. Lastly, the important feature of area is calculated using (9), where a and b represent the number of pixels within the shape. An ROI vector is constructed, consisting of a ROI at position A and a ROI at position B [13], [14]. These GLCM texture features contribute to the overall classification performance of the GLCM-CNN model. 𝐸 = 𝑎 𝑏 (7) 𝑃 = (𝑃𝑎 ,𝑏, 𝐴 𝑒𝑑𝑔𝑒[𝑃] = 𝑎, 𝐵 𝑒𝑑𝑔𝑒 [𝑃] = 𝑏) (8) 𝐴 = (𝐴𝑎,𝑏 , 𝐴 𝑅𝑂𝐼[𝐴𝑟𝑒𝑎] = 𝑎, 𝐵 𝑅𝑂𝐼 [𝐴𝑟𝑒𝑎] = 𝑏) (9) The modified technique depicted in Figure 4 is referred to as GLCM-CNN, with the architecture consisting of sub-sampling layers and three types of layers: convolutional layers followed by an output layer. The purpose of using the GLCM is to extract features that characterize a set of structures, this helps decrease the misclassification of cancerous tumors in images. In (10) outlines the sub-sampling function, wherein the input patch window function denotes the neighborhood X(n, n), and calculates the average value of this neighborhood. The output layer is directly linked to the last convolutional layer and comprises four to five neurons, each representing various stages of cancer. This layer produces the ultimate output utilizing th e acquired features, thereby facilitating the classification process. 𝐴𝑗 = (𝐴𝑗 𝑛∗𝑛 ∗ 𝑋(𝑛, 𝑛)) 𝑛∗𝑛 𝑎𝑣𝑔 (10) Figure 4. Architecture of GLCM-CNN During CNN training, the objective is to minimize the average squared error obtained from the training data. In (11) defines the average squared error, where X represents the total number of training data samples, d_j(n) denotes the target of the actual class, B stands for the number of batches, y_j(n) signifies the output of the jth layer, and E_average indicates the normalized value of the squared error. Eaverage = 1 X ∑ (dj (n) − yj (n))2 B a=1 (11)
- 6. ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091 3088 Step 4: Active contour method is utilized for refining the identification of liver tumors [15] Anil and Dayananda [11] presented a comprehensive convex active contour model (ACM) for reflex image segmentation to refine the boundaries obtained in step 3 of our methodology. This technique involves three main steps: contour Initialization, regional term formulation, and boundary term formulation. The ACM is utilized to enhance and fine-tune the tumor boundaries.It aims to recover lost borders or eliminate unwanted borders from the image. The function hr is a region function that evaluates the exterior and interior regions based on specific constraints.The function u operates on the image domain and assigns a value between 0 and 1 to each pixel position x in the image. For a detailed explanation of this step, please refer to [11]. 𝑚𝑖𝑛0 ≤𝑢≤1(∫ 𝑔𝑏 |∇𝑢|𝑑𝑥 + 𝜆 ∫ ℎ𝑟𝑢𝑑𝑥 𝛺 𝛺 ) (12) 4. RESULTS AND DISCUSSION In this section, we conducted a comprehensive comparison and evaluation of various segmentation techniques vis-à-vis our proposed method. Through meticulous analysis and comparison with other methodologies [16]‒[19], we have established that our segmentation approach outperforms these methods, exhibiting a notable improvement in dice precision ranging from 6% to 10%. In (13) delineates the accuracy calculation. It's important to note that the specific equations and intricacies regarding accuracy computation may vary depending on the methodology elucidated in the original papers [16]‒[19]. For a more thorough understanding, we recommend referring to the original research papers for a detailed explanation of the accuracy calculation methodology. A = No.of correct identification total number of test data 𝑋 100 (13) Table 2 also features the maximum and minimum values for each category of tumors. It's crucial to emphasize that without access to the actual table and specific data, we unable to provide exact scores and comparisons for each metric. Figures 5 and 6 provides a quantitative evaluation of the proposed segmentation technique, comparing it to CNN and GLCM-CNN methods. Analysis from the table reveals that the GLCM-CNN method achieves higher accuracy. Table 3 shows the classification precision of two models: CNN and GLCM-CNN. Table 4 further presents an assessment of the proposed segmentation procedure, analyzing segmentation results for small, medium, and large tumors. Metrics such as relative volume differences (RVD), dice per case, volume overlap error (VOE), average symmetric difference (ASD), and maximum symmetric difference (MSD) are utilized to quantify the scores [20]‒[22]. Nevertheless, this table facilitates a thorough assessment of the segmentation results, offering valuable insights into the performance of the proposed segmentation technique [23]‒[25] across various tumor sizes. For detailed information and precise values. Table 2. Segmentation results comparison with other approaches Method/Author Dice per case ASD (mm) RVD VOE MSD (mm) Chris et al. [16] 0.56 - - - - Yuan et al. [17] 0.56 1.151 0.228 0.378 6.269 Chlebus et al. [18] 0.65 11.11 0.380 0.625 16.71 Chen et al. [6] 0.67 6.36 0.41 0.564 11.69 Proposedmethod[CNN+GLCM] 0.86 4.753 0.234 0.312 10.21 Figure 5. Graphical representation of classification precision of CNN and GLCM-CNN Figure 6. Graphical illustration of scores (Table 4)
- 7. Int J Artif Intell ISSN: 2252-8938 Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil) 3089 Table 3. Classification precision of CNN and GLCM-CNN Iterations CNN GLCM-CNN H E C 25 66.78 52.16 79.65 88.69 50 77.25 55.24 86.63 88.62 100 86.34 69.23 93.65 89.72 200 89.89 89.91 94.65 91.76 Table 4. Evaluation of the proposed tumor segmentation technique quantitatively DICE VOE (%) RVD (%) ASD (mm) MSD (mm) Small tumor 0.87 ± 2.25 36.91 ± 8.26 24.69 ± 5.36 0.76 ± 0.21 0.98 ± 0.30 Max 0.87 44.65 29.31 0.91 0.92 Min 0.71 24.12 20.62 0.54 0.69 Medium tumor 0.81 ± 0.25 37.62 ± 9.62. 22.54 ± 3.72 0.73 ± 0.51 1.14 ± 0.52 Max 0.91 46.03 25.56 0.85 0.86 Min 0.73 32.45 19.62 0.56 0.68 Large tumor 0.91 ± 1.25 38.95 ± 3.62 19.23 ± 2.56 0.78 ± 0.32 1.12 ± 0.84 Max 0.94 41.53 21.79 0.92 0.85 Min 0.75 35.31 17.23 0.52 0.67 Average 0.87 ± 7.06 28.05 ± 4.32 23.25 ± 2.95 0.77 ± 0.56 1.11 ± 0.21 5. CONCLUSION Liver cancer, the deadliest form globally, demands precise analysis and swift tumor identification across all stages to improve survival rates. A new method, using CT scan images, is being developed to detect liver cancer stages efficiently. Initial steps involve image preprocessing, converting to grayscale and applying median filtering. A novel segmentation technique, including CNN, accurately segments cancerous polyps. Key properties like eccentricity, perimeter, and area are extracted using the GLCM method from the segmented ROI. The ACM refines tumor localization. This combined approach has shown promising results, achieving a dice similarity coefficient of 81.8 ± 14.5% during training and 80.6 ± 12.9% during testing on CT images from 132 patients. APPENDIX Table 1. Different stages of cancer Stage Stage grouping Stage description 1 C0 C0 indicates No evidence of a primary tumor. 1A C1a L0 D0 C1a indicates not grown to blood vessels and it is only tumor size varies from 1-2 cm. L0 indicates has not spread to nearby lymph nodes. D0 indicates has not spread to distant sites. 1B C1b L0 D0 C1b indicates not grown to blood vessels and it is only tumor size varies from 2-4 cm. L0 indicates has not spread to nearby lymph nodes. D0 indicates has not spread to distant sites. 2 C2 L0 D0 C2 indicates may be growth of single tumor, size is great than 2 cm and parallelly has grown into blood vessels or may be grown of more than one tumor but not larger than 6cm. L0 indicates has not spread to nearby lymph nodes. D0 indicates has not spread to distant sites. 3A C3 L0 D0 C3 indicates growth of more than one tumor, size is great than 6 cm L0 indicates has not spread to nearby lymph nodes. D0 indicates has not spread to distant sites. 3B C4 L0 D0 C4 indicates one tumorof somewhat size that has grown-up intoa mainbranchof a large vein of the. L0 indicates has not spread to nearby lymph nodes. D0 indicates has not spread to distant sites. 4A Any C L1 D0 L1 indicates growthof single tumour ormultipletumors ofanysize has blowout toneighboringlymph nodes. D0 indicates has not blowout to detached sites. 4B Any C Any L D1 D1 indicates It has spread to detached organs such as the bones or lungs. REFERENCES [1] E. Smith, “Worldcancerday 2017: livercancer, a global challenge thanks toviruses and alcohol,” Cancer Research UK, 2017, Accessed: Jun. 10, 2018. [Online]. Available: https://news.cancerresearchuk.org/2017/02/13/world-cancer-day-2017-liver-cancer- a-global-challenge-thanks-to-viruses-and-alcohol/ [2] N. Nanda, P. Kakkar, andS. Nagpal, “Computer-aidedsegmentationof liver lesions in CT scans using cascaded convolutional
- 8. ISSN: 2252-8938 Int J Artif Intell, Vol. 13, No. 3, September 2024: 3083-3091 3090 neural networks andgeneticallyoptimisedclassifier,”ArabianJournal for Scienceand Engineering, vol. 44, no. 4, pp. 4049– 4062, 2019, doi: 10.1007/s13369-019-03735-8. [3] R. Dey and Y. Hong, “Hybrid cascaded neural network for liver lesion segmentation,” in 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), 2020, pp. 1173–1177, doi: 10.1109/ISBI45749.2020.9098656. [4] S. I. TurabandV. K. Kadam, “Livercancer detection and grading using efficient computer vision techniques,” Annals of the Romanian Society for Cell Biology, vol. 25, no. 2, pp. 1740–1755, 2021. [5] A. Krishan andD. Mittal,“Ensembledlivercancer detection andclassification usingCT images,” Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine, vol. 235, no. 2, pp. 232-244, 2021, doi: 10.1177/0954411920971888. [6] F. Chen, J. Liu, Z. Zhao, M. Zhu, andH. Liao, “Three-dimensional feature-enhancednetworkforautomatic femur segmentation,” IEEE Journal of Biomedical andHealth Informatics, vol. 23, no. 1, pp. 243–252, 2019, doi: 10.1109/JBHI.2017.2785389. [7] S. Almotairi,G. Kareem, M.Aouf, B. Almutairi, andM. A. M. Salem, “Liver tumor segmentation in CT scans using modified segnet,” Sensors, vol. 20, no. 5, 2020, doi: 10.3390/s20051516. [8] C. A. Hamm et al., “Deeplearningfor livertumordiagnosis part I: development of a convolutional neural network classifier for multi-phasic MRI,” European Radiology, vol. 29, no. 7, pp. 3338–3347, 2019. [9] C. J. Wanget al., “Deeplearningfor liver tumordiagnosis part II: convolutionalneural network interpretation using radiologic imaging features,” European Radiology, vol. 29, no. 7, pp. 3348–3357, 2019, doi: 10.1007/s00330-019-06214-8. [10] S. A. Azer, “Deep learningwith convolutional neural networks foridentificationof liver masses andhepatocellular carcinoma: A systematic review,” World Journal of Gastrointestinal Oncology, vol. 11, no. 12, pp. 1218–1230, 2019, doi: 10.4251/wjgo.v11.i12.1218. [11] B. C. Anil andP. Dayananda, “Automatic livertumorsegmentation basedon multi-level deep convolutional networks and fractal residual network,” IETEJournal of Research, vol. 69, no. 4, pp. 1925–1933, 2023, doi: 10.1080/03772063.2021.1878066. [12] U. Hameed, M. U. Rehman, A. Rehman, R. Damaševičius, A. Sattar, and T. Saba, “A deep learning approach for liver cancer detectionin CT scans,” Computer Methods in Biomechanics and Biomedical Engineering: Imaging& Visualization, vol. 11, no. 7, 2024, doi: 10.1080/21681163.2023.2280558. [13] X. Jia andM. Q. H. Meng, “A deep convolutional neural networkforbleedingdetection in wireless capsule endoscopy images,” in 2016 38thAnnual International Conferenceof the IEEE Engineeringin Medicine andBiology Society (EMBC), 2016, pp. 639– 642, doi: 10.1109/EMBC.2016.7590783. [14] K. Suzuki, J. Shiraishi, H. Abe, H. MacMahon, K. Doi, “False-positive reduction in computer-aided diagnostic scheme for detectingnodules in chest radiographs by means ofmassivetrainingartificial neuralnetwork,” AcademicRadiology, vol. 12, no. 2, pp. 191–201, 2005, doi: 10.1016/j.acra.2004.11.017. [15] M. Kass, A. Witkin,andD. Terzopoulos, “Snakes: active contour models,” International Journal of Computer Vision, vol. 1, no. 4, pp. 321–331, 1988, doi: 10.1007/BF00133570. [16] P. F. Christ et al., “Automatic liver andtumorsegmentation ofCT andMRI volumes using cascaded fully convolutional neural networks,” Arxiv-Computer Science, vol. 1, pp. 1–20, 2017. [17] Y. Yuan, “Hierarchical convolutional-deconvolutional neural networks for automatic liver and tumor segmentation,” Arxiv- Computer Science, pp. 1–4, 2017. [18] G. Chlebus, H. Meine, J. H. Moltz, andA. Schenk, “Neural network based automatic liver tumor segmentation with random forest-based candidate ltering,” Arxiv-Computer Science, vol. 1, no. 1–4, 2017. [19] X. Han, “Automatic liverlesionsegmentationusinga deep convolutional neural networkmethod,” Arxiv-Computer Science, vol. 1, pp. 1–4, 2017. [20] K. He, X. Zhang, S. Ren, andJ. Sun, “Deep residual learningforimage recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778, doi: 10.1109/CVPR.2016.90. [21] C. Peng, X. Zhang, G. Yu, G. Luo, andJ. Sun, “Large kernel matters--improve semantic segmentation by global convolutional network,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4353–4361. [22] P. Isola, J.-Y. Zhu, T.Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1125–1134. [23] M. M. Santoni, D. I. Sensuse, A. M. Arymurthy,and M. I. Fanany, “Cattle race classification using gray level co -occurrence matrix convolutional neural networks,” Procedia Computer Science, vol. 59, pp. 493–502, 2015, doi: 10.1016/j.procs.2015.07.525. [24] S. S. Roy, S. Roy, P. Mukherjee,andA. H. Roy, “An automated liver tumour segmentation and classification model by deep learningbasedapproaches,” Computer Methods in Biomechanics andBiomedical Engineering:Imaging& Visualization, vol. 11, no. 3, pp. 638-650, 2023, doi: 10.1080/21681163.2022.2099300. [25] M. Hussain,N. Saher, andS. Qadri, “Computer vision approachfor liver tumor classificationusingCT dataset,” Applied Artificial Intelligence, vol. 36, no. 1, 2022, doi: 10.1080/08839514.2022.2055395. BIOGRAPHIES OF AUTHORS Bellary Chiterki Anil is working as Associate Professor & Head, Department of Computer Science and Engineering (AI & ML), JSS Academy of Technical Education, Bengaluru. He has completed B.E. degree in CSE Engineering (VTU) from RYMEC, Bellary, Karnataka, M.Tech. in CSE from SBMJCE, Karnataka, and Ph.D. in CSE Engineering from VTU, Belagavi in 2021. He has a teaching experience of more than 10 years. His area of researchworks are data mining and image processing. He has published over 25 papers in international journals and conferences. He is a member of CSI, ISTE societies in India. He is a reviewer for International Journals and conferences. He can be contacted at email: [email protected].
- 9. Int J Artif Intell ISSN: 2252-8938 Identifying liver cancer cellsusing cascaded convolutional neural network … (Bellary Chiterki Anil) 3091 Arun Kumar Gowdru is working as Professor & HOD, Department of ECE, JSSATE, NOIDA, UP. He has completed Diploma in E&C Engineering from Bapuji Polytechnic, Davanagere, B.E. degree in ECE (VTU) from STJIT, Ranebennur, Karnataka, M.Tech. in Digital Communications & Networking from Govt. UBDTCE, Davanagere, Karnataka, and Ph.D. in ECE from VTU, Belagavi in 2016. He has a teaching experience of more than 16 years. He can be contacted at email: [email protected]. Dayananda Pruthviraja is currently working as professor & HOD in the Department of IT at MITB. He Obtained PhD degree fromVTU and MTech degree from RVCE. His focus area is image processing & information retrieval. He was with MSRIT, Bengaluru, India. He has published many papers in national and international journals in the field of image processing and retrieval. He can be contacted at email: [email protected]. Niranjan C. Kundur holds a Doctorate of Computer and Information Sciences from VTU University, India in 2023. He is currently an associate professor at Department of Computer Science in JSS ATEB, VTU University, India. His research includes computer vision, pattern recognition, machine learning, data mining, deep learning, and artifical intelligence. He has published over 30 papers in international journals and conferences. He is a member of IAENG and ISTE societies in India. He is a reviewer for international journals and conferences. He can be contacted at email: [email protected] or [email protected]. Balakrishnan Ramadoss received the M.Tech. degree in CSE in 1995 from the IIT, Delhi and the Ph.D. degree in Applied Mathematics in 1983 from IIT, Bombay. Currently, he is working as a Professor (HAG) Computer Applications at NIT, Tiruchirappalli. He has 30+ years of teaching & research experience. Under his guidance, 13 have successfully completed Ph.D. programme. He has 80+ research publications in SCI/SCIE/Scopus and reputed international conferences. His research interests include security and privacy in big data and cloud, software testing, and information retrieval. He can be contacted at email: [email protected].