A computer vision approach for the classification of multi liver tumor using computed tomography scan

Abstract

This study describes the potential of computer vision (CV) approaches for liver tumor classification. Two-dimensional (2D) computed Tomography (CT) images dataset of benign and malignant liver tumors that were cholangiocarcinoma, focal nodular hyperplasia, hepatic adenoma, hemangioma, hepatoblastoma, and hepatocellular carcinoma was acquired for this study. The CT dataset comprising 150 images, each sized at (512 × 512), encompassing various types of liver tumors. This dataset consisted of a total of 900 (150 × 6) CT images representing six benign and malignant liver tumor types. To enhance data quality, a Mean filter was applied for noise reduction, followed by the selection of two regions of interest (ROIs) from each liver image. Subsequently, the preprocessed data was subjected to feature extraction, resulting in 67 multi-features per image, incorporating histogram, spectral, and texture features. From these features, 21 optimized multi-features were derived through the implementation of a correlation-based feature selection (CFS) algorithm. These optimized multi-features formed the basis for analysis and were fed into six classifiers: multilayer perceptron (MLP), logistic regression, random subspace, decision tree, produce error reduction, and multiclass classifier. The performance evaluation of these classifiers was conducted using 10-fold cross-validation techniques. The MLP showed a better accuracy of 97.67% on the optimized feature dataset among all the deployed CV classifiers. The experimental findings indicated that the suggested approach was systematic and resilient, offering valuable assistance to radiologists in detecting liver tumor diseases through CT Dataset images, even amid differing imaging standards.

Introduction

In this study, the main focus is on Computed Tomography (CT) scans, classification, liver tumors, multilayer perceptron (MLP), and optimized multi-features. These keywords highlight the central aspects of the research and guide readers toward the methodological and clinical relevance of the work. Liver tumors played a significant part in the history of liver diseases and the remarkable progress in medicine. Liver cancer is one of the greatest liver diseases and the leading cause of cancer-related deaths worldwide [1]. According to the Global Cancer Observatory (GLOBOCAN 2022), liver cancer accounted for more than 900,000 new cases and approximately 830,000 deaths worldwide, ranking as the third leading cause of cancer-related mortality. The incidence is particularly high in Asia and sub-Saharan Africa, where hepatitis B and C infections remain prevalent. These updated figures highlight the urgent need for improved diagnostic and classification tools. Previously, liver tumors were written down haphazardly. By analyzing these reports, doctors can spot diseases and estimate the chances of illnesses to give suitable treatments [2]. In the 20th century, important research proved that cirrhosis can result in hepatocellular carcinoma (HCC), the main type of primary liver cancer. After this breakthrough, researchers could better understand different liver diseases. The ability to use medical imaging has made a big difference in the way clinicians find and diagnose liver tumors. The most popular way to image the liver for medical purposes is computed tomography (CT) imaging. Segments of tumors in the patient’s liver are needed for many clinical applications using CT [3]. Ultrasonography, CT scans, and MRI make detecting and assessing liver lesions simpler. Thanks to new technologies, experts can now divide liver tumors into two categories, identify these as benign or malignant, and list hepatocellular carcinoma as an example of malignancy.

There are a variety of liver tumors, and managing each one depends on its type. HCC, which is the most widely found liver cancer, is caused mainly by hepatocytes. Because chronic liver diseases raise the chances of HCC, regularly searching for this condition is essential. In the early phases, HCC usually does not give any signs of disease, so diagnosis is possible only through tools such as ultrasonography, CT scans, MRI, and measuring alpha-fetoprotein [4]. Intrahepatic Cholangiocarcinoma (ICC) matters greatly because it develops within the bile ducts inside the liver. Usually, ICC lesions appear as areas of decreased attenuation and an outer ring of low enhancement initially, gradually becoming denser toward the center later. When completing a diagnosis, health experts rely on images and sometimes conduct a biopsy. At the same time, the place where ICC grows and its late discovery make it hard to manage with standard treatments [5]. If detected early, hepatoblastoma in young children can be dealt with successfully by using surgery, chemotherapy, and, in some instances, a liver transplant. Hemangiomas made of blood vessels and adenomas that happen because of contraceptives make up many of the benign types of liver tumors doctors see. Usually, no symptoms are seen from Focal Nodular Hyperplasia (FNH), a noncancerous growth that seldom requires any medical treatment. Because each benign tumor is different, it is vital to group them correctly for proper treatment. The area within the liver where metastatic tumors form is affected by the primary cancer’s original location. Tumors from both liver and bile duct cells are challenging to recognize and manage since they have unique characteristics [6]. How liver tumors are staged and graded is necessary to check for cancer spread, look at cell features, and steer the course of treatment. Better surgical procedures, increased use of liver transplantations, and newer drugs have come into use as liver cancer treatments have changed [7]. Recognizing liver tumors quickly is necessary since the probability of recovery often relies on their stage. Scientists are dedicated to exploring how liver tumors take shape to discover early indicators and effective therapies [8]. They indicate that medicine specialists are determined to improve liver tumor care constantly.

Computer vision (CV) plays a significant role in helping to recognize and evaluate liver tumors. Using computational algorithms on medical information and images helps boost the objectivity, efficiency, and accuracy of classifying liver tumors. Using CV techniques helps detect, estimate results, and develop therapeutic plans. ML helps identify liver tumors by itself and locate them precisely in CT scans [9]. By automating the identification process, radiologists can identify unusual results faster, which makes diagnostics more efficient and requires less time spent by radiologists. CV algorithms can mark liver tumors by locating their boundaries in medical imaging which helps to evaluate their characteristics and boost the accuracy of classifying and staging them. Such algorithms can detect important aspects of images that might be hard for humans to see, for example, the texture, shape, and intensity of the image. CV combines imaging techniques that support more accurate classifications and show a complete picture of tumor features [10].

This study [11] described an approach to divide images of the cancer-affected liver using spatial fuzzy clustering. After dividing the cancerous parts, key features were collected, and the regions were grouped into two categories of cancer. Both C4.5 and MLP classifiers were tested and got accurate results: 95.02% for the C4.5 and 89.15% for the MLP. The authors tried to predict the early recurrence of HCC by applying ML methods and analyzing images of hematoxylin and eosin-stained tissues. The research included 158 patients diagnosed with HCC and meeting Milan criteria who had surgery. People in Group 1 got HCC after their first resection, Group 2 got cancer again approximately one to two years later, and Group 3 showed no sign of HCC within four years of their resection. An SVM prediction model was used to split the groups, and the accuracy was 89.9%. The work [12] proposed a novel automated system that integrates three distinct algorithms. Specifically, GA is used for SVM optimization, SVM is used for classification, and linear discriminant analysis (LDA) is used for dimensionality reduction. The three models were combined to produce one black box model named LDA-GA-SVM. Improvements in the accuracy of HCC prediction are demonstrated by experimental findings on a publicly accessible HCC dataset. With the suggested approach, 90.30% accuracy was attained.

The work [13] proposed a study to build HCC classification techniques for predicting the existence of HCC, and different machine learning (ML) techniques (linear regression algorithm, reduce pruning error tree, classification) were used. The total accuracy of the proposed model was between 93.2% and 95.6%. The work [14] developed a machine learning (ML) model for predicting overall survival (OS) and proposing initial therapy options in HCC. A study was provided in work [15] for CT liver tumor segmentation. The study suggested an enhanced segmentation technique based on the fast fuzzy c-mean clustering technique (FFCM). The liver CT images’ contrast is improved by adjusting intensity values and removing high frequencies through the median filter method and histogram equalization. FFCM is used to separate the liver tumors from the segmented liver. Six distinct indicators are used in a quantitative analysis to evaluate segmentation results. Overall, 95% accuracy was achieved from the proposed approach. The work [16] proposed study in which the platform’s primary objective was to create a superior pathological learning dataset that enables increased accessibility. The challenge’s goal was to assess both new and current methods for automatically detecting liver cancer in whole-slide photos. In the challenge, participants had to assess how well automated algorithms performed on two distinct tasks using analytical data and statistical measures. The work [17] proposed an approach that uses the multi-scale Gabor rotation-invariant local binary pattern (MGRLBP) to obtain the final texture feature of a biomedical image. To access the classification results, both qualitative and quantitative techniques were applied. Using the proposed MSBP and MGRLBP techniques, the simulation was done with MATLAB2013a environment. The suggested approach generates more accurate texture characteristics for medical image data. Efficiency of over 93% was attained by applying the MGRLBP and MSBP methods. The work [18] recommended a study focusing on ML approaches with several automated Regions of Interest (ROI) for liver tumor classification. Dataset images were converted to grayscale, and then histogram equalization was applied to enhance the contrast. Using 165 cases, the work [19] suggests a novel machine learning method for identifying HCC. The normalizing strategy is applied during the preprocessing stage. Two applications of the evolutionary algorithm and stratified 5-fold cross-validation technique is made: one for feature selection and the other for parameter optimization. Table 1 summarize the literature review section.

Table 1.

Summary of related work on liver tumor classification

Study	Methodology	Dataset / Modality	Accuracy
Das et al. (2019) [11]	Modified fuzzy clustering + DT	CT images	95%
Saito et al. (2021) [20]	ML on histopathology images	158 HCC patients	89.9%
Ali et al. (2021) [12]	LDA-GA-SVM hybrid model	Public HCC dataset	90.3%
Hashem et al. (2020) [13]	Linear regression, classification	HCC patients with HCV	95.6%
Anter et al. (2019) [15]	Hybrid segmentation	CT images	95%
Proposed Study	Multi-feature extraction + CFS + MLP	6 tumor types	97.67%

Several denoising systems have been applied in medical imaging to enhance the quality of CT scans before feature extraction and classification. Traditional approaches, such as Mean and Median filters, have been widely used to suppress random noise, but they often blur fine edges and critical tumor boundaries. Advanced methods such as Wiener filtering and Wavelet-based denoising have shown better preservation of image details while reducing noise [14–16]. More recently, anisotropic diffusion filtering and Non-Local Means (NLM) algorithms have been successfully applied to liver CT scans, as they are able to reduce speckle and Gaussian noise without significantly compromising structural information. In the current study, we incorporated Mean and Kuwahara filters, which effectively improved image clarity and contrast, thereby enhancing the reliability of the extracted multi-features. This choice was motivated by their computational simplicity and robustness, making them well-suited for clinical environments where time efficiency is crucial. By situating our approach within the context of existing denoising techniques, it is evident that careful preprocessing plays a pivotal role in achieving high classification accuracy in liver tumor detection.

The main findings of this study demonstrate that the proposed computer vision–based framework achieved an overall classification accuracy of 97.67% for distinguishing six types of benign and malignant liver tumors using CT images. This remarkable success is primarily due to.

• Applying effective denoising filters (Mean and Kuwahara) to improve image quality.

• Extracting a comprehensive set of histogram, texture, and spectral features.

• Optimizing the feature set through Correlation-Based Feature Selection (CFS) to ensure computational efficiency and higher accuracy. Among the classifiers evaluated, the Multilayer Perceptron (MLP) outperformed all others, confirming its robustness for this medical application.

The novelty of this work lies in combining traditional denoising techniques with optimized multi-feature selection and applying them systematically to multi-class liver tumor classification. Unlike many previous studies that focused only on binary classification (e.g., HCC vs. non-HCC) or achieved lower accuracy levels with conventional classifiers, this study provides a generalized framework capable of classifying six different tumor types with high reliability. The contribution of this research is twofold: first, it establishes an efficient preprocessing-to-classification pipeline that enhances diagnostic accuracy; second, it demonstrates the potential of computer vision tools to assist radiologists by providing a reliable decision-support system. This contribution strengthens the integration of artificial intelligence into medical imaging and opens pathways for extending the approach to other imaging modalities such as MRI and ultrasound.

Materials and methods

This study comprises six CT images dataset of benign and malignant liver tumors, namely, Cholangiocarcinoma (CC), Focal Nodular Hyperplasia (FNH), Hepatic Adenoma (HA), Hemangioma (HG), Hepatoblastoma (HB), and HCC as shown in Fig. 1.

Fig. 1.

Sample of Liver tumors CT-scan image dataset

A conventional 2D CT dataset consisting of 150 CT images, each with a size of 512512 pixels, for six category of liver tumor that is acquired from publicly available data repository Radiopaedia [21]. Consequently, 900 (1506) CT scan images datasets included both malignant and benign liver tumors. An experienced radiologist meticulously analyzed The CT liver dataset using a range of medical tests and biopsy reports. This approach was arduous and time-consuming, resulting in significant expenses.

Proposed methodology

The proposed methodology has following steps. Firstly, we acquired CT scan images and subsequently resized them to a dimension of 512512 pixels. The CT color images are converted to grayscale images in the initial phase. Following this, the second step applied a Mean filter with a mask size of 3 to the grayscale image. The grayscale image was converted to a Natural Binary code image in the third step. Subsequently, the fourth step applied a Kurwahara filter [22] with a mask size of 3 to the Natural Binary code image. The final step involved the application of the Gray Level Quantization segmentation [23] technique with a threshold value of 4. Following the image processing steps, a multi-feature dataset was extracted, encompassing histogram features, texture features, and Spectrum features, with two regions of interest (ROI) identified for each image. The multi-feature dataset underwent preprocessing, and feature optimization was executed through a correlation-based feature selection technique, identifying 21 optimized features. Finally, computer vision-based classifiers were deployed on the optimized features to achieve accurate classification as shown in Fig. 2.

Fig. 2.

Proposed Methodology Framework

Image preprocessing

The 2D CT liver dataset was initially converted into grayscale. To address non-uniformities and enrich contrast during data acquisition, a mean filter technique was employed. However, speckle noise, stemming from environmental factors affecting the imaging sensor, was still apparent. To mitigate this, a Kurwahara filter was applied to reduce the speckle noise and further enhance contrast. Noisy pixel values were then substituted with their respective average values to refine the data. As a result, a significantly enhanced and smooth grayscale CT liver dataset was obtained. After this, taking two non-overlapping (ROIs) on each image using (CVIP) software version 5.7 h [24] and stored in gray level 8-bit (.bmp) format.

Automated natural binary Gray level quantization segmentation

This method in image processing is used to organize and identify the different levels of light and dark in a picture. Grouping pixels as a few grey levels lowers the level of detail in the image. Converting grayscale values in an image makes it easier to study and process with segmentation methods [23]. First, the original image’s pixels are organized into several levels. In quantization, unique intensity values are reduced, producing a simpler representation in grayscale. Therefore, every pixel in the image receives a particular intensity level based on its original one. It provides a basis for further division into segments, allowing for spotting regions with comparable properties. For image analysis tasks such as detecting objects, grey-level quantization segmentation helps because it simplifies texture differences and improves the results given by later algorithms [25]. It is essential in medical imaging because it supports recognizing and separating different entities, such as tumors, in the pictures. Deciding on the grade of quantization plays a vital role in managing the struggle between saving image content and reducing the amount of work for segmentation.

First, the process breaks the original image’s pixel intensities into several levels. When quantization is done, the range of different intensity values is decreased, giving a simpler look to the grayscale image. Hence, every pixel in the image gets a specific amount of detail based on how bright it was initially. It allows for progression to segmentation, which helps identify regions with similar output levels. Image analysis tasks, such as detecting objects, benefit significantly from gray-level quantization segmentation since it increases the following algorithms’ efficiency and precision by making the image variations less complicated [25]. Besides, it helps medical imaging by identifying and outlining things in the images, like tumors or anatomical structures. How many quantization levels are used impacts how the image’s details are preserved and how easy the segmentation process becomes.

Multi-Feature dataset acquisition

In this study, liver tumor CT images were utilized to extract a variety of multi-features, encompassing texture, spectral, and histogram characteristics. These features were computed as follows: five for each first and second-order texture, including five mean values oriented in four directions (0°, 45°, 90°, 135°). Additionally, twenty-eight binary features were derived from regions of interest (ROIs) with a width and height of 10 pixels, alongside seven RST features and six spectral attributes, incorporating an extra accumulative mean value.

Consequently, a total of 57 features were extracted for each ROI, culminating in a dataset comprising 95,760 features (1680 × 57). This dataset was generated across various ROI sizes. All experimentation was conducted using the Computer Vision and Image Processing (CVIP) software version 5.9 h, operating on an HP^® Core i7 processor clocked at 2.6 GHz (GHz), with 8 gigabytes (GB) of RAM, and running on a 64-bit Windows-10 operating system. The segmentation framework proposed in Fig. 3 delineates the methodology employed in this research.

Fig. 3.

Proposed segmentation approach outcomes

Histogram features

These characteristics, which are based on the intensity of individual pixels, are sometimes called first-order statistical features or histogram features [26]. Their properties are defined by (1).

1

In this case, U(T) depicts the grayscale values, and R represents the total number of pixels in Eq. (1). In Eqs. 2–6, several histogram characteristics are computed and shown. Equation (2) now displays the mean feature. With “r” standing for gray level values and “n” and “o” for rows and columns, we get Eq. (2).

2

In Eq. (2), The gray level values are shown by “r” while “n” and “o” represents the rows and columns. The contrast of the image was described by using the Standard deviation (SD), and shown in Eq. (3).

3

Equation (3) shows the results of describing the image’s contrast using the standard deviation (SD). Skew asymmetry is assessed when there is no symmetry around the central pixel value and is computed using Eq. (4).

4

An equation describing the distribution of grayscale values is given by the symbol “energy” (5).

5

The “entropy” of the picture is depicted in Eq. (6), which describes its unpredictability.

6

Texture features

These features, also referred to as second-order statistical features [27], capture the visual characteristics of the object and are assessed using Gray-Level Co-occurrence Metrics (GLCM). Computed across four directions—0°, 45°, 90°, and 135°—up to a five-pixel distance, these five features include inverse difference, entropy, correlation, inertia, and energy. Mathematically, the “energy” feature is defined by Eq. (7).

7

The correlation approach is used to describe the degree to which two pixels are similar at a given distance.

8

The overall content of the image is shown by the entropy. The result may be seen in Eq. (9).

9

Equation (10) evaluates the inverse difference as a probability of a locally homogeneous.

10

Equation (11) evaluates the technique of inertia, which is defined by the contrast.

11

Spectral features

These features rely on pixel frequency values [28], commonly employed in texture-oriented image classification tasks. The energy delineates various regions or areas, known as sectors and rings.

12

Feature optimization

Feature selection involves extracting many features to identify the most relevant ones. This often involves managing large datasets, presenting a considerable challenge. The essential task is to minimize the dimension of the feature space, facilitating efficient differentiation between classes. Various strategies are applied to pinpoint the most distinguishing characteristics [29]. These pivotal features are then used to achieve classification accuracy that is both cost-effective and efficient. The selected ones transform a new space with lower dimensionality to optimize features. The objective is to retain the original data structure as much as possible. This feature optimization not only decreases execution time and costs but also yields results almost equivalent to those obtained in the original feature space.

In the context of liver tumor classification, it was observed that not all collected features were meaningful. Managing the extensive dataset, comprising 118,800 (661800) features vector space (FVS), posed a significant challenge. The need to curtail the number of features in the space was evident. The correlation-based feature selection (CFS) technique [30] was employed to achieve this task. The CFS utilizes information theory principles, explicitly incorporating the concept of entropy as expressed in Eq. (13).

13

The variable W is described in Eq. (14)

14

Z P is the main probability, whereas P () is the secondary probability. Equation (15) displayed the supplementary data W.

15

Equation (16) describes the typical association between characteristics “W” and “Z”.

16

Where Eq. (17) expresses the connection among features as symmetrical uncertainty (SU):

17

In order to demonstrate the relationship between continuous and discrete features, the typically used features allow for entropy-supported quantification. The original feature space was used to extract 21 optimized features using CFS. In Table 2, we can see the optimized feature space.

Table 2.

Selected optimized multi features for liver tumor classification

Features
1. Inertia Average	12.SpotWave_Standard_Deviation
2. LevelLevel_Mean	13.WaveRipple_Mean
3. LevelLevel_Standard_Deviation	14.WaveRipple_Standard_Deviation
4. LevelEdge_Standard_Deviation	15.RippleRipple_Standard_Deviation
5. LevelSpot_Mean	16.Spectral_DC
6. LevelSpot_Standard_Deviation	17.Ring1
7. LevelRipple_Mean	18.Sector1
8. EdgeEdge_Standard_Deviation	19.Sector2
9. Edgespot_Mean	20.Sector4
10. EdgeRipple_Mean	21.Sector5
11. EdgeRipple_Standard_Deviation

Ultimately, the initial dataset of 118,800 (66 × 1800) features was condensed to a CFS-based dataset of 37,800 (21 × 1800) features for each size of ROIs pertaining to liver tumors. This reduced feature dataset was then utilized for classification purposes.

Classification

In the study, six computer vision (CV) classifiers namely, Multilayer Perceptron (MLP), Logistic (Lg), Multiclass Classifier (MCC), Random Subspace (RS), Decision Tree (DT) and JRip (JR) were utilized on the liver tumor multi feature dataset using the WEKA tool, version 3.8.6 [31]. The MLP classifier, as described in work [32], functions by computing the weighted sum of inputs and biases using the summation function , as defined in Eq. 18. While the specific equation is not provided in the given context, it serves as the mathematical operation employed in the MLP classifier to calculate the weighted sum of inputs and biases.

18

In the provided equation, ‘k’ represents the number of inputs, where denotes the input variable ‘I’, represents the bias term, and denotes the weight. Among the multiple activation functions available for Multilayer Perceptron (MLP), one is provided below:

19

The output of neuron j can be computed as follows:

20

Table 3 outlines the parameter settings for the MLP classifier, while Fig. 4 illustrates the statistical multi-feature analysis MLP framework, encompassing all the regulatory parameters [33].

Table 3.

Deployed MLP classifier parameter values

Parameter	Value
InputLayers	1
HiddenLayers	15
LearningRate	0.4
Momentum	0.1
Epochs	315
ValidationThreshold	20
Neurons	15

Fig. 4.

MLP framework for liver tumor classification using optimized statistical multi-features

The “green” color denotes the first layer of the MLP framework, which comprises 21 features in the input layer. The second layer, depicted in “red,” represents the hidden layer, consisting of 15 neurons. The third layer, highlighted in “yellow,” corresponds to the output layer and includes six nodes representing the weights of the hidden layers. Above the framework, the regulatory parameters and their respective values are displayed.

Results and discussion

In this study, six computer vision (CV) classifiers namely, Multilayer Perceptron (MLP), Logistic (Lg), Multiclass Classifier (MCC), Random Subspace (RS), Decision Tree (DT) and JRip (JR) were utilized for the classification of six types of liver tumors based on CT scan images. At the initial stage of experimentation, we extracted texture, histogram and spectral features, by combining all these features we generated multi features dataset. The Lg, DT, MCC, JR, RS and MLP shows classification accuracy 95%, 95.39%, 95.44%, 95.89%, 97.39% and 97.50% respectively without deploying feature selection technique. It was observed that MLP showed better accuracy result of 97.5% among employed classifiers. The high success rate of 97.67% achieved in this study can be attributed to three key factors: effective preprocessing, optimized feature selection, and robust classification. By employing Mean and Kuwahara filters, the dataset was significantly denoised and enhanced, which improved the clarity of tumor boundaries. The application of Correlation-Based Feature Selection (CFS) further refined the dataset by reducing redundant features, allowing only the most discriminative attributes to contribute to classification. Finally, the use of the Multilayer Perceptron (MLP) provided superior learning ability for complex nonlinear relationships, outperforming other classifiers in this research. When compared with similar studies, such as those using C4.5 (95.02%), SVM-based systems (90.30%), and hybrid fuzzy clustering techniques (95%), the proposed approach demonstrates a clear improvement in accuracy. This superiority highlights the effectiveness of combining classical denoising filters with feature optimization and advanced machine learning classifiers. Thus, the proposed framework not only improves classification performance but also establishes a reliable and efficient tool for assisting radiologists in clinical decision-making. The result of MLP with other CV classifiers with other performance monitoring factors like kappa statistics, true positive (TP), false positive (FP), receiver-operating characteristic (ROC), recall, mean absolute error (MAE), root mean squared error (RMSE), time, and precision were described in Table 4.

Table 4.

The liver tumor classification table based on CT scan multi-feature dataset

Classifiers	Kappa Statistics	TP Rate	FP Rate	ROC	Recall	MAE	RMSE	Time (Sec)	Precision
Lg	0.94	0.950	0.010	0.993	0.950	0.0171	0.1266	3.47	0.950
DT	0.9447	0.954	0.009	0.991	0.954	0.0647	0.133	0.59	0.957
MCC	0.9513	0.959	0.008	0.933	0.959	0.0197	0.1166	18.72	0.960
JR	0.9507	0.959	0.008	0.989	0.959	0.019	0.1138	0.67	0.959
RS	0.9687	0.974	0.005	0.995	0.974	0.0199	0.0881	0.25	0.974
MLP	0.97	0.975	0.005	0.985	0.975	0.0097	0.0892	47.01	0.975

Figure 5 displays the outcomes of the CV classifier that was put into action on a dataset with several features. The MLP classifier outperformed the other classifiers used on the multi-feature dataset, with a classification accuracy of 97.5%.

Fig. 5.

The implemented CV classifier’s results on multi-features dataset

In the second phase of experimentation, we deployed features selection technique and obtained optimized multi features dataset, and then applied same CV classifiers on the optimized features dataset, it has been observed that these classifiers performed well on the optimized dataset. The DT, Lg, MCC, JR, RS, and MLP shows classification accuracy 95.83%, 96.83%, 96.28%, 97%, 97.44% and 97.67% respectively that shown in Table 5.

Table 5.

The liver tumor classification table based on CT scan optimized multi-feature dataset

Classifiers	Kappa Statistics	TP Rate	FP Rate	ROC	Recall	MAE	RMSE	Time (Sec)	Precision
DT	0.95	0.958	0.008	0.991	0.958	0.0627	0.1304	0.19	0.960
Lg	0.9553	0.963	0.007	0.993	0.963	0.0151	0.1044	2.94	0.963
MCC	0.962	0.968	0.006	0.995	0.968	0.0158	0.0947	0.98	0.968
JR	0.964	0.970	0.006	0.988	0.970	0.0155	0.0984	0.18	0.970
RS	0.9693	0.974	0.005	0.996	0.974	0.0201	0.0894	0.09	0.975
MLP	0.972	0.977	0.005	0.986	0.977	0.0116	0.0869	3.91	0.977

In Fig. 6, we can see all the CV classifier’s results on the optimized multi-feature dataset. The MLP classifier outperformed the other deployed classifiers on the optimized multi-feature dataset, with a classification accuracy of 97.67%.

Fig. 6.

The implemented CV classifier’s results on multi-features dataset

Likewise, the results of the MLP classifier on the optimized multi-feature dataset are displayed in Table 6. The instances located on the diagonal represent data correctly classified into their respective classes, while instances elsewhere denote misclassifications. It was observed that the MLP classifier demonstrated superior accuracy compared to other deployed classifiers.

Table 6.

Confusion matrix for liver tumor hybrid multi-feature dataset using MLP

Classified as	CC	FNH	HCC	HG	HA	HB	Total
Cholangiocarcinoma (CC)	288	5	4	0	0	3	300
Focal Nodular Hyperplasia (FNH)	5	295	0	0	0	0	300
Hepatocellular Carcinoma (HCC)	4	0	292	0	0	4	300
Hemangioma (HG)	0	0	0	297	3	0	300
Hepatic Adenoma (HA)	0	0	0	3	297	0	300
Hepatoblastoma (HB)	6	1	4	0	0	289	300

The findings of classifying six different types of liver tumors (both benign and malignant) using a CT dataset were as follows: cholangiocarcinoma (96%), hepatoblastoma (97.33%), hepatocellular carcinoma (99%), focal nodular hyperplasia (99%), hemangioma (96.3%), and hepatic adenoma (99%). Figure 7 shows the graphical accuracy results of six different CT datasets about liver tumors that were processed using MLP classifiers on an optimized features dataset.

Fig. 7.

The Accuracy Graph Using MLP Classifier on Optimized multifeatured dataset

In addition to the accuracy results, further evaluations can provide deeper insights into the robustness of the proposed framework. For example, analyzing sensitivity, specificity, F1-score, and area under the ROC curve (AUC) would offer a more comprehensive view of classifier performance. Such evaluations highlight not only how well the model classifies each tumor type but also how reliably it distinguishes between malignant and benign cases. Moreover, comparative runtime analysis and confusion matrix visualization may demonstrate the computational efficiency and consistency of the proposed method. These additional evaluations would strengthen the reliability of the findings and help validate the practical applicability of the system in clinical settings. Figure 8 presents a comparative graph showcasing the classification accuracy of six liver tumor CT scan datasets using both a multi-featured dataset and an optimized multi-features dataset. This graph provides a comprehensive overview of the overall accuracy results achieved by the deployed CV classifiers on liver CT images. It is evident that the accuracy results were notably superior with the optimized multi-features dataset compared to the multi-featured dataset.

Fig. 8.

The Accuracy Graph Comparison Using multi features and optimized multi features datasets

Conclusion

This research aimed to classify six liver tumor images CT datasets through multi-feature analysis. Six machine learning classifiers, including MLP, Lg, MCC, RS, DT, and JR, were utilized with an optimized multi-feature dataset. In addition to overall classification accuracy, this dataset was utilized to assess various other quality measuring parameters, as discussed earlier in the results and discussion section. The results obtained from the deployed classifiers were satisfactory, with the MLP classifier exhibiting exceptionally high performance compared to others. Specifically, the MLP classifier yielded an impressive overall accuracy of 97.67% when classifying both Benign and Malignant liver tumors. It is noteworthy to mention that the optimization of the multi-feature dataset was crucial for achieving such remarkable accuracy results. This optimization was achieved through the application of supervised correlation-based feature selection (CFS) technique, which effectively reduced the dimensionality of the overall feature space. Without this optimization step, it would not have been feasible to attain such high accuracy results within a short execution time frame. Therefore, the employment of the CFS technique played a pivotal role in enhancing the efficiency and effectiveness of the classification process for liver tumor images CT datasets.

Future work

In future research, we will explore the impact of variations in texture feature values under different illumination factors using deep learning algorithms. Additionally, we aim to investigate the integration of different modalities such as MRI and ultrasound data within the framework of deep learning. This research will enhance our understanding of how deep learning models can adapt to variations in texture features and leverage the complementary information from diverse imaging modalities for improved liver tumor classification and diagnosis.

Author contributions

M.Z.: Conceptualization, Methodology, Data curation, Investigation, Writing – original draft. J.Z.: Project administration, Resources, Validation, Supervision. W.K.M.: Data curation, Validation, Writing – review & editing. F.K.K.: Data curation, Writing – review & editing. S.M.M.: Formal analysis, Funding acquisition, Investigation, Validation. and M.A.: Data curation, Writing – review & editing.

Funding

This research was supported by the Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2025R300), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Data availability

The Liver tumor CT scan dataset accessed from publicly available database Radiopaedia: (https://radiopaedia.org/). The multi feature datasets extracted and analyzed during the current study available from the corresponding author on reasonable request.

Declarations

Ethical approval

Not applicable.

Consent for publication

All the authors agreed to be published.

Competing interests

The authors declare no competing interests.