







































Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 
67-75 

67 

 

 

 

Article 

Hybrid contrast-limited adaptive histogram 

equalization and Deep Learning techniques for 

improving  liver tumor detection  
Priyah. R, S. Kamalakkannan*  

Vels Institute of Science, Technology & Advanced Studies, Chennai, Tamil Nadu, India 

A R T I C L E   I N F O 
 

Article history: 
Received 07 April 2025  
Received in revised form 
22 May 2025 
Accepted 31 May 2025 
 
Keywords:  
Contrast-limited adaptive histogram equalization 
(CLAHE), Convolutional neural networks(CNN), 
Generative adversarial networks(GAN’s),  
Transfer learning, Liver tumor detection,  
Medical images 
 
*Corresponding author 
Email address: 
kannan.scs@vistas.ac.in 
 
 
DOI: 10.55670/fpll.futech.4.3.7 

A B S T R A C T 
 

Deep Learning and advanced image processing can enhance the detection and 
prognosis of liver cancer using medical imaging, such as Magnetic Resonance 
Imaging (MRI) and Computed Tomography (CT) scans. Liver cancer detection 
is a challenging task due to factors such as poor contrast, noise in imaging 
techniques, limited annotated datasets, and the complex characteristics of 
tumors. This study proposes a hybrid technique that combines Contrast-
Limited Adaptive Histogram Equalization (CLAHE), Convolutional Neural 
Networks (CNNs), Generative Adversarial Networks (GANs), and Transfer 
Learning (TL) to improve the precision and accuracy of liver tumor detection. 
A conventional technique for image enhancement, CLAHE increases the 
contrast of medical images, making malignant tumors more apparent. CLAHE, 
however, does not provide a thorough tumor characterization; instead, it 
focuses on enhancing image quality. CNN is used to extract features, find and 
learn important patterns, such as edges, textures, and shapes that are 
pertinent to the diagnosis of tumors.  Finally, TL utilizes pre-trained models 
(Inception V3) for classification, enabling the effective learning of tumor 
features and achieving high diagnostic precision with fewer computational 
resources. A hybrid approach combining CNN, GAN, and TL may give an 
integrated and effective solution for identifying and diagnosing liver tumors. 
The hybrid technique performed significantly better than independent DL 
approaches, achieving an accuracy of 93.3%, a sensitivity of 92.2%, a 
specificity of 94.5%, and an F1-score of 92.8%. 

 

1. Introduction 

Liver cancer is one of the main causes of mortality for 
individuals all over the world. It is challenging and time-
consuming to manually identify the cancer tissue in the 
present scenario. The American Cancer Society predicted 
that 611,720 people would die from cancer in 2024, while 
2,001,140 new cases would be diagnosed. Lung, prostate, 
and colorectal cancers are the leading causes of mortality for 
men, and lung, breast, and colorectal cancers are the leading 
causes for women, accounting for an estimated 611,720 
deaths, or about 1671 deaths per day. These findings 
highlight the persistent difficulties that cancer presents and 
the need for continual research and advancement in cancer 
diagnosis and treatment [1]. The cells of the liver are the 
source of liver cancer, which can either originate in the liver 
directly (primary liver cancer) or migrate from other regions 
of the body to the liver (secondary liver cancer). Major risk 
factors for primary liver cancer include cirrhosis, chronic 
liver disorders such as hepatitis B or C infection, and heavy 

 

use of alcohol. Imaging techniques, including ultrasound, CT, 
MRI, and PET scans, as well as blood testing for tumor 
markers such as alpha-fetoprotein (AFP), are necessary for 
its diagnosis. By detecting liver masses, nodules, or lesions, 
these imaging methods help patients with liver cancer with 
diagnosis, staging, and therapy planning [2]. Techniques for 
image enhancement are essential for raising the visual 
quality and clinical utility of CT  images. Improvements in 
image enhancement technologies are now crucial for 
medical practitioners to correctly interpret CT scans, as 
medical imaging remains an essential component of 
diagnostic and treatment planning. Tumor segmentation in 
CT images of the liver can be used to determine the severity 
of the tumor, schedule therapies, predict outcomes, and 
monitor clinical response [3]. DL techniques, such as CNN, 
have shown considerable promise in addressing complex 
medical imaging challenges, including organ segmentation. 
CNNs are ideal for medical image analysis because they can 

May 2025| Volume 01 | Issue 01 | Pages 01- 03 

ISSN 2832-0379 

Open Access Journal 

Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.0.0.0 

 

 

Future Technology 

 

Future Technology 

ISSN 2832-0379 

August 2025| Volume 04 | Issue 03 | Pages 67-75 

Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.4.3.7 

 

 

 

 

 

Open Access Journal 

mailto:kannan.scs@vistas.ac.in
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.0.0.0
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.4.3.7


Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

68 

 

automatically detect complicated patterns and discriminate 
between tissues and organs with high accuracy [4]. 

 

 

 

 

 

 

 

 

 

 

 

 

 

1.1 Medical imaging techniques  
Computed Tomography(CT) is a basic imaging method 

that is widely accessible.  A CT scan is a diagnostic technique 
that uses computer processing to create incredibly precise 
cross-sectional images, or slices, of the bones, blood vessels, 
and soft tissues.  Compared to standard X-rays, CT scan 
images offer more information and enable a more accurate 
assessment of numerous anatomical elements. This imaging 
approach detects irregularities in soft tissues, circulatory 
systems, and bones, enabling a more comprehensive 
examination. It can detect small tumors and provide detailed 
imaging of liver architecture. CT imaging is utilized to 
identify tumors, cysts, and masses in the liver, including 
metastases, cholangiocarcinoma, and hepatocellular 
carcinoma (HCC). It differentiates between cancerous and 
benign tumors and assesses tumor vascularity utilizing 
enhancement trends, such as arterial enhancement for HCC. 
CT images help in liver cancer staging to assess the location, 
size, and quantity of tumors and detect blood vessel 
invasion, such as portal vein thrombosis. Then directs 
interventional treatments such as transarterial 
chemoembolization (TACE), radiofrequency ablation (RFA), 
or surgical excision or transplantation [5]. In liver imaging, 
CT is a crucial tool, especially when combined with advanced 
procedures such as triphasic scanning. Depending on clinical 
needs, CT is often combined with other imaging modalities, 
such as MRI or ultrasound, to achieve the most effective 
results. 

 Magnetic Resonance Imaging (MRI) is an essential tool 
for detecting liver cancer is magnetic resonance imaging, or 
MRI. It is generally accepted to be the most sensitive imaging 
technique for assessing the liver in individuals with chronic 
liver disease and describing liver cancers. High contrasts 
between soft tissues in the acquired images are used by MRI, 
a precise and accurate technique for tumor diagnosis. 
Although this feature makes MRI especially useful for 
detecting and describing cancers, circumstances pertaining 
to the patient as well as the operator may affect the 
diagnostic efficacy of MRI. Patient-related factors, such as 
claustrophobia, implanted materials or devices, and 
unpleasant circumstances, may restrict the use of MRI and 
affect the caliber of the findings. Because MRI is non-
invasive, has excellent imaging capabilities, and can provide 

both anatomic and physiological information, it is the most 
commonly used method for diagnosing and characterizing 
liver cancer. It is particularly important for individuals who 
are at high risk and who need a precise medical diagnosis 
and planning. 

1.2 Contrast-limited adaptive histogram equalization  
A more complex method called histogram equalization 

modifies an image's dynamic range by changing the pixel 
values in accordance with the image's intensity histogram. 
CLAHE is a prevalent image preprocessing technique that 
enhances contrast by altering image histograms, especially 
in darker places. It has been demonstrated that CLAHE helps 
to improve image clarity, especially in medical imaging, 
where precise tissue segmentation necessitates a high 
contrast between tissues. While CLAHE enhances contrast, it 
is unable to fully address other quality issues, including 
noise, uneven illumination, and border sharpness that are 
critical for precise kidney segmentation. CNN algorithms 
must employ sophisticated or relevant preprocessing 
techniques to improve the quality of images and achieve the 
best segmentation accuracy [6]. CLAHE is frequently used in 
medical imaging, particularly for liver image enhancement, 
to enhance the visibility of features and structures in low-
contrast areas. This method is especially useful for 
enhancing the clarity of liver examination modalities, such as 
CT, MRI, and ultrasound, where subtle differences in pixel 
intensity can make it challenging to detect specific 
abnormalities. It facilitates the differentiation between 
normal and pathological areas by highlighting subtle 
variations in tissue architecture, blood vessels, and liver 
lesions. In technologies such as ultrasound, the contrast-
limiting approach lowers noise amplification, which is 
critical. It is ideal for photographs with poor illumination or 
shifting contrast, as it focuses on emphasizing specific areas 
of interest while preserving the entire image [7]. 

1.3 Deep Learning methods 
CNNs have been shown to be highly efficient in medical 

imaging applications, such as detecting liver cancer. They 
serve as an important tool for detecting liver cancers such as 
HCC and its metastases because of their ability to 
autonomously capture and retrieve structured data from 
medical images. CNNs can learn complex structures and 
patterns using unprocessed visual data, eliminating the need 
for human feature engineering. CNNs are particularly 
excellent at finding minor anomalies that traditional 
methods may ignore, as well as distinguishing between 
malignant and benign liver cancers. It can be used with 
histopathology slides, CT, MRI, ultrasound, and other 
imaging modalities. CNNs use a single pipeline to combine 
preprocessing, feature extraction, classification, and 
prediction.  

A technique for detecting liver cancer is transfer 
learning, particularly when there is a dearth of labeled 
medical imaging data. It entails using pre-trained models on 
big datasets (like Inception V3) and optimizing them on 
datasets related to liver cancer imaging in order to improve 
identification, classification, and segmentation tasks. Large 
annotated datasets are not as necessary with transfer 
learning, which also speeds up model convergence and 
enhances performance. TL greatly raises the precision and 
effectiveness of liver tumor detection tasks by utilizing the 
advantages of pre-trained models. 
The objectives of the study are as follows: 

Abbreviations 

AES  AutoEncoders 

CNN  Convolutional Neural Networks 

Ct  Computed Tomography 

DBN  Deep Belief Networks  

FN  False Negatives 

FP  False positives 

LSTM  Long Short Term Memory  

LiTs  Liver Tumour Segmentation 

MRI  Magnetic Resonance Imaging 

PET  Positron Emission Tomography 

RCNN  Recurrent Neural Networks 

TCIA  The Cancer Imaging Archive 

TP  True Positives 

TN  True Negatives 



Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

69 

 

• To improve the identification and categorization of liver 
cancers in medical imaging by creating a hybrid 
framework that combines CLAHE for image 
enhancement and CNN feature extraction  

• To apply the pre-trained TL models for the classification 
of liver cancer diseases to substantially improve the 
diagnostic accuracy of liver cancer. 
The paper's subsequent sections are arranged as 

follows: Section 2 provides an overview of the current 
literature related to the present study. A thorough 
explanation of the techniques utilized to arrive at the best 
results for the detection and categorization of liver cancer 
may be found in Section 3. Section 4 presents the outcome 
and accompanying comments, and Section 5 concludes the 
study with recommendations for future research. 

2. Related works  

The incidence of liver cancer has been continuously 
increasing over time, making it a serious worldwide health 
problem. For prompt treatment and better patient outcomes, 
liver cancer must be accurately detected and classified.  
Sajjanar et al. [8] investigate a variety of strategies, including 
feature extraction using traditional machine learning (ML) 
algorithms and segmentation using deep learning (DL) 
models. The use of ensemble approaches, which aggregate 
predictions from multiple models to enhance segmentation 
accuracy, has also been explored. Huang et al. [9] provide a 
detailed review of a revolutionary image-enhancing 
procedure that highlights an image's important features by 
adaptively adjusting the brightness and contrast. Then, to 
improve the accuracy of tumor region detection, a  DL based 
segmentation network was provided. This network was 
explicitly trained using the upgraded pictures. Kaur et al. 
[10] conduct a thorough examination and comparison of 
numerous image enhancement methods used on medical 
pictures, with a focus on CT imaging of the abdomen. The 
study analyzes the benefits and drawbacks of each algorithm 
by categorizing it into three distinct categories: DL, 
transformation, and histogram-based methods. The research 
conducts an in-depth investigation, taking into account 
factors such as the Structural Similarity Index (SSIM), Mean 
Squared Error (MSE), Average Mean Brightness Error 
(AMBE), Entropy, and Peak Signal-to-Noise Ratio (PSNR), to 
determine how well different techniques perform. Rani et al. 
[11] suggest a method for classifying and segmenting liver 
tumors automatically. The three primary parts of our 
suggested architecture are a pixel-wise classification unit for 
detecting liver anomalies, a preprocessing unit to improve 
image contrast, and a Masked Recurrent CNN(RCNN) for 
liver segmentation. Researchers obtain Dice similarity 
coefficients of 96% for liver segmentation and 98% for 
lesion detection with the MICCAI'2027 liver tumor 
segmentation (LITS) database. 

Wu et al. [12] reviewed the latest advancements in DL 
methods used for liver cancer multimodal fusion image 
segmentation. The use of DL in multimodal image 
segmentation for liver cancer is revolutionizing medical 
imaging and is anticipated to improve the precision and 
effectiveness. This review offers medical professionals 
helpful advice and insights. Singh et al. [13] examined DL 
models for automated liver and tumor area segmentation 
from CT scans to enhance liver cancer treatment planning 
and diagnostic precision, utilizing 3200 CT images by 
choosing around 70% for testing, while the remaining 30% 
were used for training. Preprocessing techniques include 
Histogram Equalization and Auto Windowing HU 

(Hounsfield Units), which help preserve the images' 
improved clarity. The CNN models, such as ResUNet, VGG16, 
and VGG19, were employed and assessed using the 
performance measures, including accuracy, Dice coefficient, 
IoU (Intersection over Union), precision, and recall. 
Compared to VGG16 at 93.5% and VGG19 at 91.5%, its 
hybrid design, ResUNet, which combines U-Net and ResNet, 
produced improved outcomes with an accuracy of 97.3%. A 
model's capacity to distinguish between healthy and 
malignant tissues can be enhanced by extracting pertinent 
information using the Differential CNN model. The Kernel 
Extreme Learning Machine (KELM) model is used to classify 
features into benign and malignant categories. The 
Differential Biogeography-Based Optimization method 
(DBBOA) finds near-optimal solutions by fine-tuning the 
parameters. The DL-based categorization model is trained 
through this tweaking procedure [14]. 

Das et al. [15] employed three distinct numerical 
mapping approaches to digitize the gene sequences. After 
digitalization, these sequences of DNA were first analyzed in 
two ways: as 2D spectrogram images and as 1D signals. 
First, the CNN model was used to analyze the digital 
sequences as a one-dimensional signal. Second, two distinct 
2D CNN models were used to analyze DNA signals after they 
were transformed into 2D spectrogram pictures. VGG16 
produced the feature vectors for the first model, and SVM 
categorized them. The second model did fine-tuning and 
added new layers to the VGG16 final output layers.  Hameed 
et al. [16] used CT images from the Kaggle dataset to develop 
a profound model for identifying liver malignant areas. CT 
scans from the dataset obtained from the Kaggle platform 
undergo basic preprocessing. An accuracy value of 94.3% 
was attained in tumor detection when the suggested model's 
performance was assessed using various performance 
criteria. 

Gedeon et al. [17] utilized  LivlesioNet, which is based 
on DenseNet, to extract features from the input. At each step, 
the model produces useful feature maps. The improved 
multi-scale convolutional layer and LivlesioNet's efforts 
have reduced the number of parameters, making training 
with a minimal dataset possible. A bridge scale (BS) is then 
suggested to combine multi-scale spatial characteristics with 
the goal of eliminating duplicate features and modifying 
feature map weights in order to increase accuracy. 
Furthermore, a fully-connected layer and a SoftMax classifier 
are coupled for additional classification following the 
concatenation layer. Bhaskar et al. [18] proposed a method 
that correctly classifies liver histopathology images as either 
malignant or non-cancerous, supporting the early 
identification of the emergence of liver cancer cells that may 
invade or disseminate. Although histopathological image 
analysis (HIA) is crucial for identifying the growth of cancer 
cells, it is laborious, prone to errors, and reliant on the skill 
of the pathologist. In order to increase precision as well as 
effectiveness in liver cancer cell proliferation, this research 
suggests an automatic HIA that makes use of DL. The model 
employs multiple instances learning for picture-level 
classification, ResNet50 for patch-level feature extraction, 
OpenCV libraries for image preprocessing, and whole slide 
image (WSI) input. Ma et al. [19] constructed a Transformer 
structure block with a dense residual attention CenterNet 
network to suggest a liver tumor detection technique called 
TDCenterNet. A dense residual attention network is 
intended to improve feature flow in cases where lesion 
features are not sufficiently extracted. The dependencies 
between global characteristics are captured by embedding 



Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

70 

 

transformer structural blocks. To retrieve lesion 
characteristics of varying sizes, atrous spatial pyramid 
pooling is incorporated. A knowledge distillation training 
approach is created to enhance TDCenterNet's performance. 
According to earlier research investigations, the most 
difficult jobs are those involving computationally complex, 
sensitive parameter setups, misdetection, and 
misclassification. 

3. Methodology   

The proposed approach combines CLAHE for image 
enhancement, CNN for feature extraction, and transfer 
learning for classification, as shown in Figure 1. The 
suggested hybrid method's deep learning component 
includes: CLAHE, CNN,  Data Augmentation, and Transfer 
Learning models. The detailed flow diagram is shown in 
Figure 2. 

 
Figure 1. Proposed method 

 

 

 

 

 

 

 
 
 
 
 
 
 
 
 
 
 

 

Figure 2. Detailed Workflow –Proposed Method 

3.1 CLAHE 
A histogram equalization method called CLAHE 

prevents noise amplification and over-enhancement by 
minimizing amplification in homogenous areas, thereby 
improving image contrast. The suggested approach applies 
CLAHE to liver imaging pre-processing data, highlighting 
tumor areas and enhancing fine detail visibility. The image 
has been split into tiles or blocks of a predetermined size 
(e.g., 8x8 pixels) that do not overlap. Each tile undergoes 
histogram equalization to improve contrast by redistributing 
pixel intensity levels. A clip limit, often known as a contrast 
limit, is used to avoid over-enhancing noise and artifacts. To 
preserve a smooth contrast, additional histogram counts are 
uniformly allocated. Then the borders of neighboring tiles 
are blended using bilinear interpolation, ensuring seamless 
transitions across the entire image. The input images are 
optimized for further DL analysis because of this 
preprocessing phase.  The DL model uses the CLAHE-

enhanced images as inputs. Robust tumor detection is 
achieved by the hybrid approach, which combines the 
advantages of both conventional and AI-driven approaches 
through improved contrast and sophisticated feature 
extraction. Figure 3 shows the images of liver tumors before 
and after CLAHE application. Tumor borders are more 
obvious on the right side, which shows the heightened 
contrast produced by CLAHE, whereas the left side shows 
the original picture with low contrast. 

 
Figure 3. Sample liver tumour image after and before CLAHE 

3.2 Convolutional neural networks (CNNs) 
The basic components for image categorization 

problems nowadays are CNNs. However, extracting 
pertinent information from a picture is another extremely 
valuable task that is carried out before classification is 
performed. CNNs use feature extraction to identify 
important patterns in an image so they can categorize it.  The 
workflow for extracting features from images using CNN is 
shown in Figure 4. The Conv2d layer is the primary 
component of a CNN. It extracts key characteristics and 
reveals hidden patterns. These characteristics serve as 
building blocks that enable the network to comprehend the 
content of the image. The Conv2d layer analyzes the picture 
using filters, also known as kernels. These filters examine 
tiny regions of the image at a time as they move across it. It 
converts the raw pixels into useful visualizations and 
retrieves pertinent information as it goes. This method can 
identify edges, forms, and their significant aspects in the 
image. To extract features, a specific CNN architecture is 
created.  The learnt characteristics are captured in feature 
maps created by the Conv2D layer. The resulting feature 
maps are sent to the next layers (such as dense and pooling) 
for additional processing. A variety of liver imaging datasets 
with tumor areas identified are used to train the algorithm 
[20,21]. 

3.3 Data augmentation 
Data augmentation is the process of applying several 

alterations to the current data in order to artificially extend 
it. Current data points are changed into additional instances 
of data with the same semantic labels. GANs and 
AutoEncoders (AEs) are used in this study to generate the 
images.  GAN uses two neural networks: the discriminator 
and the generator. While the discriminator's role is to 
distinguish between generated and actual data, the 
generator's goal is to create data that is indistinguishable 
from actual data. This configuration allows you to produce 
highly realistic data samples.  AEs, on the other hand, receive 
a reduced version of the data in a smaller-dimensional latent 
space and then utilize it to recreate the original data. It is 
feasible to change the compressed form to generate fresh 
data points. These strategies enable the deliberate and 



Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

71 

 

nuanced creation of data, which has significant benefits in 
terms of increasing both the variety and accuracy of training 
datasets [22,23]. 

 

 
Figure 4. Feature Extraction using CNN 

3.4 Inception V3 
InceptionV3 is a popular deep CNN for image 

categorization applications, including medical imaging. Its 
architecture, which was designed for successful feature 
extraction, is particularly suitable for complex medical data, 
such as ultrasound, CT, or MRI scans used to predict liver 
cancer. InceptionV3 extracts and learns meaningful 
information from images by processing them across many 
layers. A number of convolutional and pooling layers are 
applied to the image. These layers extract low-level features 
from the input liver picture, such as edges, textures, and 
basic shapes. InceptionV3 is distinguished by its Inception 
modules, which have been designed to: 
• Extracting multi-scale features involves applying many 

convolutional filters in parallel, each with a different 
size (e.g., 1x1, 3x3, and 5x5). 

• Minimize the cost of computation by reducing the 
dimensionality of feature maps through the use of 1x1 
convolutions. 

• Identify both large lesions and tiny cancers by 
examining the picture at various dimensions. 
The size, form, and texture of liver tumors might vary. 

Tumor identification is improved by the Inception modules' 
parallel convolutions, which guarantee that features of 
various scales are recorded. This enables the network to 
focus on the most significant characteristics, including 
aberrant texturing or tumor borders. Auxiliary classifiers are 
incorporated into intermediate layers of InceptionV3 to aid 
with learning and avoid overfitting. This is especially helpful 

in cases where the dataset is minimal, such in the diagnosis 
of liver cancer. InceptionV3 utilizes global average pooling at 
the network's end, rather than fully connected layers, to 
mitigate overfitting and enhance the model's capacity for 
generalization [24]. 

4. Results and findings  

A hybrid approach integrating CLAHE, CNN, GAN, and 
Inception v3 for liver tumor classification is implemented in 
Python using the TCIA dataset.   

4.1 Dataset description  
The Liver Tumor Collection is maintained by TCIA (The 

Cancer Imaging Archive). It is a database of imaging results 
for liver cancer, together with related clinical data. TCIA is a 
sizable collection of cancer medical images that are openly 
accessible. According to a common illness picture modality 
or kind or study emphasis, the imaging data are arranged as 
"collections." TCIA's main file format for radiological imaging 
is DICOM. The provision of image-related supporting 
information, such as patient outcomes, therapy specifics, 
genomics, and expert evaluations, is prioritized. This dataset 
includes CT imaging investigations taken before and after 
TACE  in 105 confirmed HCC patients. It contains hand-
selected semi-automated segmentations of the liver, tumor, 
and blood arteries.  

4.2 Performance analysis  
The hybrid model suggested was assessed through a 5-

fold cross-validation and a 70:30 train-test split to assess the 
robustness and reliability of the performance assessment 
methods. Final metrics were averaged across the folds 
during training with cross-validation, to allow consistency 
with results from the independent test set. Table 1 
summarizes the performance comparison. The proposed 
hybrid approach demonstrated superior performance across 
all metrics, validating its effectiveness in liver tumor 
detection. The proposed framework demonstrates notable 
improvements in classification accuracy, sensitivity, 
specificity, and resilience across various imaging settings 
when evaluated on publicly accessible liver imaging datasets 
[25].  
Accuracy: The percentage of cases (both tumor and non-
tumor) that were correctly categorized against the total 
number of cases. It shows the model's overall performance; 
however, it might not be enough when working with 

datasets that are unbalanced. 

                                   (1) 

 
Specificity: The proportion of real non-tumor instances that 
were accurately classified as such. It assesses the model's 
capacity to prevent false alarms, which occur when benign 
tissues are mistakenly identified as cancers, especially 
crucial in medical diagnostics to minimize needless 

procedures. 
 

 (2) 

Sensitivity: The proportion of real tumor cases that were 
accurately classified as tumors. It demonstrates the model's 
accuracy in tumor detection. Early diagnosis and treatment 
depend on fewer missed instances, which is ensured by high 
sensitivity. 

Sensitivity =
TP

TP+FN
                                          (3) 



Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

72 

 

F1-Score: It balances false positives and false negatives. 
When datasets are unbalanced and both false positives and 
false negatives are significant, it can be helpful. 

   
                   (4) 

W 
where 
 

 
         (5) 

                                   
Comparisons were made between the suggested hybrid  
approach and standalone DL models with and without 
CLAHE. 

Table 1.  DL Methods Vs Metrics(without CLAHE) 

Methods 
Vs  

Measures 

Accuracy 
(%) 

Specificity 
(%) 

Sensitivity 
(%) 

F1-
Score 
(%) 

CNN -
Inception 

V3 

86.7 84.2 87.6 86.3 

RNN 79.5 77.6 81.3 80.4 

LSTM 83.1 81.2 85.0 83.5 

DBN 80.4 78.8 81.9 81.1 

 

From Table  1, it is clear that all DL models perform lower 
without CLAHE, most likely because there is less contrast 
enhancement, making it more difficult to see faint tumor 
characteristics in the input images. It is evident from Table 2 
that CNN performs better than the other models.  With an 
accuracy of 93.9.3%, sensitivity of 92.2%, specificity of 
94.5%, and F1-Score of 92.8, the hybrid approach 
significantly outperformed solo approaches. The integration 
of CNN, CLAHE, and InceptionV3   contributed to improved 
detection and reduced false positive rates. 

Table 2  DL Methods Vs Metrics(with CLAHE) 

Methods Vs 
Measures 

Accuracy 
(%) 

Specificity 
(%) 

Sensitivit
y (%) 

F1-Score 
(%) 

CNN -
Inception V3 

93.9 92.2 94.7 92.8 

RNN 87.5 83.5 87.1 86.3 

LSTM 92.7 88.6 91.7 90.9 

DBN 90.3 86.2 89.5 88.3 

 

Figure 5 shows the accuracy (%) of several models 
(CNN -Inception V3, RNN, LSTM, and DBN) in detecting liver 
tumors with and without the use of CLAHE for image 
enhancement. After applying CLAHE, all models exhibit a 
notable increase in accuracy, as seen by the larger red levels 
relative to the blue levels. When compared to the other 
models, RNN exhibits the lowest accuracy (79.5%) without 
CLAHE and a slight improvement (87.5%) with CLAHE. Both 
DBN and LSTM exhibit notable improvements, with LSTM 
attaining 92.7% accuracy with CLAHE and DBN reaching 
90.3%. Figure 6 shows the specificity (%) of several liver 
tumor detection models (CNN -Inception V3, RNN, LSTM, 
and DBN) with and without the use of (CLAHE) for image 
enhancement is displayed in Figure 6. Applying CLAHE 
significantly improves the specificity of all models.  

 
Figure 5. Accuracy- Hybrid and standalone DL methods 

 

 

Figure 6. Specificity- Hybrid and standalone DL methods 

Because CLAHE increases visual contrast, models can 
more easily discern between tumor and non-tumor areas. 
CNN's exceptional ability to extract spatial characteristics 
and prevent false positives is demonstrated by its maximum 
specificity with CLAHE (92.2%). Although specificity is lower 
(84.2%) without CLAHE, it is still superior to most models. 
RNN has the lowest performance (77.6%) of any model 
without CLAHE. Although it increases to 83.5% with CLAHE, 
it is still behind CNN-Inception V3, LSTM, and DBN. With 
CLAHE, LSTM significantly improves from 81.2% (without 
CLAHE) to 88.6%. It can handle improved images more 
efficiently since it can process sequential information. DBN 
increases to 86.2% (with CLAHE) from 78.8% (without). 
Despite having a simpler design than CNN, it still gains a lot 
from CLAHE. Both with and without CLAHE, CNN -Inception 
V3 has the maximum specificity, proving its advantage in 
precisely detecting negative situations. 

Figure 7 shows the sensitivity of many DL models (CNN 
-Inception V3, RNN, LSTM, and DBN) with and without the 
use of CLAHE. CNN has a sensitivity of 94.7% with CLAHE 
and 87.6% without it. With CLAHE, sensitivity rises 
noticeably. All of the models' sensitivity is increased by using 
CLAHE. For CNN -Inception V3, the improvement is very 
noticeable. This implies that CLAHE preprocessing improves 
the quality of input data, which in turn helps DL models 
perform better. 

Figure 8 shows the F1-score (in %) of DL  models like 
CNN Inception V3, RNN, LSTM, and DBN, which contrasts 
them in two scenarios: one without CLAHE  and one with 
CLAHE. CNN demonstrates a notable improvement in F1-
score of 86.3% without CLAHE and 92.8% with CLAHE. CNN 



Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

73 

 

and LSTM models show the highest gain, indicating that 
CLAHE improves both models' performance and data quality. 
This is consistent with CLAHE's objective of enhancing 
feature clarity for tasks using images or data. This study 
proposes a novel approach to liver tumor detection using a 
hybrid framework that integrates task-specific GANs to 
generate realistic tumor patterns with CLAHE for contrast 
enhancement of liver tumor images.  Transfer Learning uses 
Inception V3, which is identified as one of the most effective 
models for medical image classification. Rather than relying 
on existing isolated techniques, the aforementioned 
components are integrated into an optimized workflow that 
respects their interdependencies. The combination of these 
integrated elements has led to a significant improvement in 
accuracy, sensitivity, and specificity with regard to correctly 
identifying liver tumors.  

 
Figure 7.  Sensitivity- Hybrid and standalone DL methods 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 
Figure 8.  F1-Score- Hybrid and standalone DL methods 

Table 3 presents a comparison table that highlights the 
positive aspects and drawbacks of the suggested approach in 
contrast to the current methods. This hybrid technique 
paves the path for more efficient and scalable medical 
imaging solutions by highlighting the complementary 
abilities of image enhancement, feature extraction, data 
augmentation, and transfer learning in addressing the 
complexity of liver tumor identification. CNN - Inception V3 
techniques have revolutionized the detection of liver cancer 
by providing automated, high-accuracy lesion detection, 
categorization, and segmentation solutions. CNN - Inception 
V3 will remain essential in enhancing liver cancer detection 
and improving patient outcomes as artificial intelligence and 
computing power continue to advance.  

 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 3. Strengths and limitations of the proposed method compared to existing methods 

Method Strengths Limitations 

Proposed Hybrid (CLAHE + GAN + 
Inception V3) 

 Improves image 
contrast with CLAHE, 
enhancing tumor 
visibility.- Task-
specific GAN generates 
realistic, tumor-like 
images.- Inception V3 
efficiently handles 
fine-grained features 
with transfer 
learning.- Integrated, 
cohesive workflow 
improves accuracy 
and consistency. 

 Computationally intensive due to multi-step processing.- Requires 
careful hyperparameter tuning for GAN and CLAHE settings.- Still 
dependent on availability of quality annotated baseline images for 
GAN training. 

CNN-based Methods 

 Learns hierarchical 
image features directly 
from data.- Flexible 
architecture 
customization. 

 Requires large annotated datasets.- Prone 
to overfitting on small datasets.- Poor 
generalization without data augmentation. 

CNN + TL 

 Utilizes pre-trained 
models, reducing 
training time and data 
requirements.- 
Improved feature 
extraction from 
medical images. 

 Limited to features learned from non-
medical (generic) images.- May not capture 
fine-grained tumor-specific patterns 
effectively. 

CNN + GAN 

 Addresses data 
scarcity by generating 
synthetic images.- 
Enhances model 
robustness with 
diverse training data. 

 GANs may produce unrealistic or noisy 
images without careful tuning.- Often lacks 
tailored tumor-specific image 
characteristics. 

 

 



Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

74 

 

Overfitting, a common machine learning issue where 
models trained on sparse data perform well on training data 
but poorly on unknown data, is what data augmentation 
attempts to minimize. CLAHE greatly improves contrast and 
the identification of critical traits in every model, resulting in 
better outcomes across every metric. CNN-Inception V3 
outperforms in all statistics due to its superior spatial 
feature extraction capabilities.  Because of its picture data 
optimization, CNN routinely outperforms RNN, LSTM, and 
DBN. Because it can handle long-term dependencies, LSTM 
beats RNN and might be beneficial for processing sequential 
image data. Because of its simplistic design and lack of 
complicated sequential or spatial feature training 
capabilities, DBN performs just marginally higher than CNN-
Inception V3. 

A revolutionary method for detecting liver cancer, 
transfer learning tackles important issues in medical 
imaging such as data variability and shortage. Better patient 
care is made possible by using pre-trained models to enable 
quicker and more accurate diagnosis. The outcomes 
validated Inception V3 high liver/tumor segmentation 
precision, which would be helpful for physicians to get 
adequate diagnostic improvement. Inception V3 
architecture, which was created especially for the medical 
imaging segmentation tasks in this study, shows promise as 
a flexible framework that might be integrated into 
automated oncological diagnostic tools and drastically alter 
the identification of liver cancer. 

 This hybrid method improves the accuracy and 
durability of liver tumor detection systems by combining the 
benefits of contrast enhancement, complex feature 
extraction, and categorization. While the hybrid technique 
produced promising results, drawbacks such as higher 
computational difficulty and reliance on high-quality 
annotations were discovered. Future study will look at 
optimization approaches and unsupervised learning to 
improve scalability and usability to larger datasets. 

5. Conclusion 

This study presented a hybrid framework that included 
CLAHE for image enhancement, CNN for feature extraction, 
and Inception v3 for tumor diagnosis. Integrating classical 
image processing with cutting-edge AI technologies provides 
an effective approach to addressing obstacles in medical 
imaging diagnosis. The results highlight the promise of 
hybrid approaches for improving detection accuracy and 
patient outcomes. CLAHE effectively increases contrast in 
liver images, making tumors more apparent and improving 
feature extraction. CNNs extract strong features by recording 
the textural and spatial information required to identify liver 
cancers. Because of its efficient design and multi-scale 
feature extraction capabilities, InceptionV3 detects liver 
cancer with high accuracy, sensitivity, specificity, and F1-
score. The hybrid framework outperforms conventional 
algorithms in terms of overall classification accuracy, F1-
score, and sensitivity. This improvement is due to an 
effective combination of DL techniques and CLAHE. The 
current work might be expanded to incorporate automatic 
liver and tumor segmentation, allowing for exact localization 
and tumor size estimation. More studies and clinical 
validation are required to turn these developments into real-
world applications. Future research could work around these 
limitations with the inclusion of automatic liver and tumor 
segmentation modules, allowing for precise tumor 
localization and size estimation for increased clinical 
applicability. The dataset may be enlarged by collaborating 

with multiple centers to regularly add a diverse cohort of 
cases to their overall model years. Furthermore,  a 
lightweight model architectures or optimization techniques 
that conserves computational expense and can be adopted in 
clinical settings, along with the auto-segmentation and the 
dataset. Generally, the framework will need longitudinal 
clinical validation and randomized trials to prove that it 
works and is safe in practice. 

Ethical issue 
The authors are aware of and comply with best practices in 

publication ethics, specifically with regard to authorship 

(avoidance of guest authorship), dual submission, 

manipulation of figures, competing interests, and compliance 

with policies on research ethics. The author adheres to 

publication requirements that the submitted work is original 

and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data 

will be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1] Siegel, R. L., Giaquinto, A. N., & Jemal, A. (2024). 

Cancer statistics, 2024. CA: a cancer journal for 

clinicians, 74(1), 12-49. 

[2]  Dumachi, A. I., & Buiu, C. (2024). Applications of 

Machine Learning in Cancer Imaging: A Review of 

Diagnostic Methods for Six Major Cancer Types. 

Electronics, 13(23), 4697. 

https://doi.org/10.3390/electronics13234697 

[3]  Tejaswi VS, Rachapudi V. Liver cancer classification 

via deep hybrid model from CT image with improved 

texture feature set and fuzzy clustering based 

segmentation. InWeb Intelligence 2024 Sep 13 (Vol. 

22, No. 3, pp. 291-314). Sage UK: London, England: 

SAGE Publications. 

[4]  Nasir, M. U., Zubair, M., Ghazal, T. M., Khan, M. F., 

Ahmad, M., Rahman, A. U., ... & Mansoor, W. (2022). 

Kidney cancer prediction empowered with 

blockchain security using transfer learning. Sensors, 

22(19), 7483. 

[5]  Wan, Y., Wang, D., Li, H., & Xu, Y. (2023). The imaging 

techniques and diagnostic performance of 

ultrasound, CT, and MRI in detecting liver steatosis 

and fat quantification: A systematic review. Journal of 

Radiation Research and Applied Sciences, 16(4), 

100658. 

[6]  Buriboev, A. S., Khashimov, A., Abduvaitov, A., & Jeon, 

H. S. (2024). CNN-Based Kidney Segmentation Using a 

Modified CLAHE Algorithm. Sensors, 24(23), 7703. 

[7]  Scherer, R., da Costa, D. R., Malek, D. A., Jammal, A. A., 

& Medeiros, F. A. (2024). Contrast-Limited Adaptive 

Histogram Equalization filters for Quality 

Enhancement for Glaucoma Fundus Imaging. 

Investigative Ophthalmology & Visual Science, 65(7), 

6476-6476. 

[8]  Sajjanar R, Dixit UD, Vagga VK. Advancements in 

hybrid approaches for brain tumor segmentation in 

MRI: a comprehensive review of machine learning 

https://doi.org/10.3390/electronics13234697


Priyah. R & S. Kamalakkannan/Future Technology                                                               August 2025| Volume 04 | Issue 03 | Pages 67-75 

75 

 

and deep learning techniques. Multimedia Tools and 

Applications. 2024 Mar;83(10):30505-39. 

[9]  Huang D, Liu Z, Li Y. Research on Tumors 

Segmentation based on Image Enhancement Method. 

arXiv preprint. 2024 Jun 7,   

https://doi.org/10.48550/arXiv.2406.05170 

[10]  H. Kaur, N. Kaur and N. Neeru, "A Comparative Study 

of Image Enhancement Algorithms for Abdomen CT 

Images," 2024 IEEE International Conference on 

Interdisciplinary Approaches in Technology and 

Management for Social Innovation (IATMSI), Gwalior, 

India, 2024, pp. 1-6, doi: 

10.1109/IATMSI60426.2024.10502768. 

[11]  Rani, R. M., Dwarakanath, B., Kathiravan, M., 

Murugesan, S., Bharathiraja, N., & Vinoth Kumar, M. 

(2024). Accurate artificial intelligence method for 

abnormality detection of CT liver images. Journal of 

Intelligent & Fuzzy Systems, (Preprint), 1-16. 

[12]  Wu, Chaopeng, Qiyao Chen, Haoyu Wang, Yu Guan, 

Zhangyang Mian, Cong Huang, Changli Ruan et al. "A 

review of deep learning approaches for multimodal 

image segmentation of liver cancer." Journal of 

Applied Clinical Medical Physics 25, no. 12 (2024): 

e14540. https://doi.org/10.1002/acm2.14540 

[13]  P. Singh, S. K. Mishra, Nidhi, J. Singh, M. Agrawal and 

P. Joshi, "Efficient Liver and Tumor Segmentation 

from CT Images Using a Hybrid ResUNet Model," 

2024 IEEE North Karnataka Subsection Flagship 

International Conference (NKCon), Bagalkote, India, 

2024, pp. 1-6, doi: 

10.1109/NKCon62728.2024.10775024 

[14]  Jesi, P. M., & Daniel, V. A. A. (2024). Differential CNN 

and KELM integration for accurate liver cancer 

detection. Biomedical Signal Processing and Control, 

95, 106419. 

[15]  Das, B., & Toraman, S. (2022). Deep transfer learning 

for automated liver cancer gene recognition using 

spectrogram images of digitized DNA sequences. 

Biomedical Signal Processing and Control, 72, 

103317. 

[16]  Hameed, U., Ur Rehman, M., Rehman, A., 

Damaševičius, R., Sattar, A., & Saba, T. (2024). A deep 

learning approach for liver cancer detection in CT 

scans. Computer Methods in Biomechanics and 

Biomedical Engineering: Imaging & Visualization, 

11(7), 2280558. 

[17]  Gedeon, K. K., & Liu, Z. (2024). Classification of liver 

lesions in CT images based on LivlesioNet, modified 

Multi-Scale CNN with bridge Scale method. 

Multimedia Tools and Applications, 83(3), 8911-

8929. 

[18]  Bhaskar, N., Kiran, J. S., Satyanarayan, S., Divya, G., 

Raju, K. S., Kanthi, M., & Patra, R. K. (2024). An 

approach for liver cancer detection from 

histopathology images using hybrid pre-trained 

models. TELKOMNIKA (Telecommunication 

Computing Electronics and Control), 22(2), 401-412. 

[19]  Ma, Ke Ouyang, Ziping Ma, Mingge Xia, Silong Xu, Ke 

Lu, Transformer dense center network for liver 

tumor detection, Biomedical Signal Processing and 

Control, Volume 91, 2024, 106066, ISSN 1746-8094, 

https://doi.org/10.1016/j.bspc.2024.106066. 

[20]  Kavitha, V. R., Jahir Hussain, F. B., Chillakuru, P., & 

Shanmugam, P. (2024). Automated Classification of 

Liver Cancer Stages Using Deep Learning on 

Histopathological Images. Traitement du Signal, 

41(1). 

[21]  Malik, M. H., Ghous, H., Rashid, T., Maryum, B., Hao, Z., 

& Umer, Q. (2024). Feature extraction-based liver 

tumor classification using Machine Learning and 

Deep Learning methods of computed tomography 

images. Cogent Engineering, 11(1), 2338994. 

[22]  Chen, Y., Lin, H., Zhang, W., Chen, W., Zhou, Z., Heidari, 

A. A., ... & Xu, G. (2024). ICycle-GAN: Improved cycle 

generative adversarial networks for liver medical 

image generation. Biomedical Signal Processing and 

Control, 92, 106100. 

[23]  Bandaru, S. C., Mohan, G. B., Kumar, R. P., & Altalbe, A. 

(2024). SwinGALE: fusion of swin transformer and 

attention mechanism for GAN-augmented liver tumor 

classification with enhanced deep learning. 

International Journal of Information Technology, 

16(8), 5351-5369 

[24]  Setiawan, W., Suhadi, M. M., & Pramudita, Y. D. 

(2024). Inception-v3 with Reduce Learning Rate for 

Optimization of Lung Cancer Histopathology 

Classification. Ingénierie des Systèmes d'Information, 

29(2). 

[25]  Makram, M., & Mohammed, A. (2024, July). Deep 

learning approach for hepatic lesion detection. In 

2024 Intelligent Methods, Systems, and Applications 

(IMSA) (pp. 312-318). IEEE. 

 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://doi.org/10.48550/arXiv.2406.05170
https://doi.org/10.1002/acm2.14540
https://doi.org/10.1016/j.bspc.2024.106066
https://creativecommons.org/licenses/by/4.0/

