Gabor Transform-Based Deep Learning System Using CNN for Melanoma Detection

Abstract

Melanoma is highly dangerous and can spread rapidly to other parts of the body. It has an increasing fatality rate among different types of cancer. Timely detection of skin malignancies can reduce overall mortality. Therefore, clinical screening methods require more time and accuracy for diagnosis. An automated, computer-aided system would facilitate earlier melanoma detection, thereby increasing patient survival rates. This paper identifies melanoma images using a Convolutional Neural Network. Skin images are preprocessed using Histogram Equalization and Gabor transforms. A Gabor filter-based Convolutional Neural Network (CNN) classifier trains and classifies the extracted features. We adopt Gabor filters because they are bandpass filters that transform a pixel into a multi-resolution kernel matrix, providing detailed information about the image. This study suggests a method with accuracy, sensitivity, and specificity of 98.58%, 98.66%, and 98.75%, respectively. This research supports SDGs 3 and 4 by facilitating early melanoma detection and enhancing AI-driven medical education.

1. Introduction

Melanoma is an atypical, uncontrolled growth of cancer cells in the skin caused by ultraviolet (UV) radiation from the sun or tanning beds, which accumulates rapidly. It is classified as non-melanoma or melanoma skin cancer. Skin cancer patients have a mortality rate of 1.4% compared to other cancer patients [1,2]. The fatality rate of skin cancer patients has experienced exponential growth since 1992. Chan. S et al. (2020) [3] discuss the prospects of rising melanoma mortality, melanoma subtypes, the need for early detection, and strategies to reduce skin cancer incidence. They also suggested that early detection of skin cancer reduces mortality more rapidly than later detection. Melanoma skin cancer is ten times more common than non-melanoma skin cancer. Melanoma survival rates depend on the stage at diagnosis and the treatment provided. According to a survey by S.M. Dawes et al. (2016) [4], white patients outnumber other skin patients, and the authors suggest that early diagnosis can lower the death rate of minority patients. Figure 1 shows a dermoscopic image of melanoma, with cancer pixels visible.

The main contributions of this research article are highlighted in the following points.

• Skin dermoscopy images are enhanced using the Histogram Equalization algorithm to improve pixel visual quality for further processing.
• The Gabor transform is applied to enhanced dermoscopy images at different orientations, phase shifts, and scales to produce Gabor dermoscopy images.
• The feature, constructed from a traditional CNN, is incorporated to classify dermoscopy skin images as melanoma or non-melanoma, using generated texture and higher-dimensional feature values from the sequence of internal layers.
• The proposed melanoma image classification approach is verified for performance using two skin-dermoscopy datasets.

2. Literature Review

In Ref. [5], Kittler et al. (2002) present a study on skin cancer detection using dermoscopy by experts and non-experts and conclude that dermoscopy, as a method, helps experienced experts detect skin cancer. In their paper [6], Gao et al. (2022) discussed the use of deep learning algorithms to improve the accuracy of medical image processing. This would help doctors find and diagnose diseases more quickly using medical images. Xu et al. (2020) [7] used the Satin Bowerbird Optimization (SBO) algorithm for skin cancer detection in clinical images. They used it at the segmentation and feature-extraction levels, achieving better results than other methods. Dildar, M. et al. (2021) [8] present a detailed review of skin cancer detection using Artificial Neural Networks (ANNs), Convolution Neural Networks (CNNs), Kohonen Neural Networks, and Generative Adversarial Neural Networks (GANs), and provide information about the datasets and open research challenges in the field. Naqvi et al. (2023) [9] discuss a study on architectures employing CNNs, VCGs, U-Net architectures, and other deep learning models to detect skin cancer. They also investigated different types of skin cancer, datasets, and resources required for deep learning methods of skin lesions. Balaha (2023) [10] and Al-Rasheed (2023) [11] propose using transfer learning to improve image processing for skin cancer detection with pre-trained CNNs. In the paper [12], L. Yu et al. (2017) used the fully convolutional residual network (FCRN) and deep neural networks for segmentation and classification, achieving first place in segmentation and second place in classification for melanoma skin cancer diagnosis on dermoscopic images. J. Daghrir et al. (2020) [13] used a combination of machine learning and deep learning to find skin cancer. They used the Morphological Snakes techniques, such as MorphACWE and MorphGAC, to segment skin lesions, and CNN models to classify them. In reference [14], S. M. et al. (2023) used the regularized extreme learning machine algorithm with a CNN to achieve an accuracy of about 98.5% in melanoma skin cancer detection. In their paper [15], S. Deivasigamani et al. (2019) applied the Gabor transform to image processing for the detection of glioma brain tumours. They converted the spatial-domain analysis into a multi-resolution image analysis and then used a feed-forward back-propagation model, achieving excellent sensitivity, specificity, positive predictive value, and accuracy. In paper [16], Keita et al. (2022) employed a kernel support vector machine (SVM) with 2D discrete wavelet transform (DWT) to classify MRI images as malignant or benign, achieving promising results in parametric evaluations. Khan M. (2019) [17] reports achieving an average accuracy of about 98% across four datasets using the 2D blue channel to enhance the contrast of skin lesions for saliency mapping and the Particle Swarm Optimization (PSO) algorithm to identify accurate borders. The work of S. Deivasigamani et al. (2016) [18] broke down the features of EEG signals using the Dual Tree Complex Wavelet Transform (DT-CWT). They then used the Adaptive Neuro-Fuzzy Inference System (ANFIS) classifier to separate the focal and non-focal EEG signals. Rachana. R. et al. produced a paper [19] based on the non-linear Optimization technique for Skin Cancer Detection by LeNet Tuning. This algorithm had a 91.99% success rate, 90.95% sensitivity, and 92.13% specificity.

The traditional algorithms described in this section have limitations for melanoma image detection systems, as outlined below.

• The traditional system used conventional CNN architectures for melanoma detection on dermoscopy images, resulting in a lower final detection rate.
• They failed to identify or differentiate dermoscopy images at low-resolution pixel formats.
• All traditional methods for melanoma image detection systems operated directly in the spatial domain. Extensive features cannot be derived from spatial dermoscopy images, thereby reducing the accuracy of melanoma image detection.
• The proposed methodologies outlined in this manuscript overcome the limitations of the traditional melanoma cancer detection system as outlined in the following points.
• The proposed method detects and classifies both low- and high-resolution dermoscopy images through an enhancement process, which overcomes the limitations of the conventional melanoma detection system.
• The proposed methodology estimated extensive feature maps from spatial dermoscopy images, which were then fed into the proposed CNN to produce melanoma classification rates, whereas the conventional CNN was fed only with spatial features.

In this way, the proposed melanoma detection system described in this paper overcomes the limitations of conventional melanoma detection methodologies.

According to the survey, CNN is a potential algorithm for image recognition. This paper proposes a Gabor-filter-based CNN for melanoma classification in dermoscopy images.

3. Deep Learning-Based Skin Cancer Detection

From the survey, early detection of melanoma increases the survival rate of patients with skin cancer. On the other hand, the melanoma death rate is increasing due to the late treatment of patients. In routine practice, skin cancer is identified using either dermoscopy or biopsy, based on the patient’s condition. This process is long, delaying the start of treatment for the patient with cancer. Therefore, computer-aided melanoma identification can deliver results to the patient more promptly. This proposed algorithm uses non-linear systematic features to differentiate healthy from diseased images. These computed non-linear systematic features are classified using a binary classifier. The accuracy of dermoscopy and biopsy reports is lower. However, with computer-assisted diagnosis, melanoma can be accurately identified more quickly. The authors used k-fold cross-validation to assess classification accuracy across different dermoscopic skin imaging datasets. The average dermoscopic image classification rate is about 96.1%.

The deep learning approach for diagnosing skin cancer (Nahata, H. 2020) [20] is gaining interest due to its high accuracy. Therefore, many research studies are using deep learning methodology for melanoma identification (J. Vineeth 2022 [21]). The deep learning method comprises several steps for processing dermoscopic lesions. An image processing system typically comprises a preprocessing stage, segmentation, feature extraction, classification, and a post-processing stage. In skin cancer detection using deep learning, preprocessing of dermoscopic images includes noise removal, resizing to match the training dataset, and contrast enhancement, while preserving hair. In the segmentation process, the preprocessed images are segmented using region-based, edge-based, or Threshold-based methods. These segmentation algorithms have merits and demerits based on the region of computation. The segmentation patch accuracy is high for these algorithms on abnormal dermoscopic images, which is a key merit of the proposed work. This work can detect and segment the outer boundary region of pixels based on abnormal dermoscopy skin images. However, it cannot detect and segment both the internal and external boundaries of the pixel region in abnormal dermoscopy skin images. This has been identified as the primary limitation of these pixel and region-based segmentation algorithms. Once the images are segmented by interest, feature extraction converts the pictorial representation into datasets using a suitable transformation. In the classification phase, the original patient datasets are compared with the trained datasets, and the output is produced, as shown in Figure 2.

3.1. Convolutional Neural Network

The Convolutional Neural Network (CNN [15]) is a widely used tool in medical image processing. It is an automated image-processing method that extracts simple features from edges and curves to analyze shapes and corners. The CNN [Sivakumar, M.S. et al. (2024) [22]] performs best at image recognition using artificial neurons within the network to process visual information. It also finds applications in segmentation and classification in medical image processing. The convolution process improves accuracy by applying the convolution function to the input and the trained data. It takes place at various stages of artificial neurons for various image characteristics. Figure 3 depicts the operations in a CNN: the convolutional blocks, in which the kernel is convolved with the pixels, thereby highlighting which features are more important across the entire dataset. This block comprises multiple layers of architecture; information from one layer passes to the next, thereby increasing precision. In the pooling layer, the information is downsampled to capture a fixed window of image data. The pooling sector has two types of functional operations based on the pooling process: Average Pooling Algorithm (APA) and Global Pooling Algorithm (GPA). The APA method places a 3 × 3 window over the features produced by the Convolution layer and then computes the average of the elements within this pooling window. The averaging process does not affect the classification results of the final pooled feature value. In the case of GPA, the maximum feature value within the pooling window is used as the pooled value. The final pooled feature value significantly affects classification results because the maximum value is selected from the Convolution layer output. Both the APA and GPA functions are detailed in the following equations.

$$Polling-APA=Average (features within the pooling window)$$

$$Polling-GPA=Maximum (features within the pooling window)$$

Both the APA and GPA performed pooling operations on the computational feature values from the Convolutional layers. The GPA method does not modify the Convolution layer output; hence, this pooling method was chosen for this study to achieve optimized classification results.

In the fully connected layer, all features from the convolutional and pooling layers are combined, and the information is processed to produce accurate results for the dermoscopy skin image recognition task.

3.2. Gabor Transform-Based Skin Cancer Identification

3.2.1. Gabor Transform

Gabor transform (Jie Yao (1995), [23]) is a specific part of short-time FFT (Fast Fourier Transform), which contains the sinusoidal and phase angle parameters. In the Gabor transform, the Gaussian window is used with a limited range of significance. As signals are represented in the time–frequency domain, when a particular signal is transformed into the time–frequency domain of another signal, it produces a similarity between the signals in that window. Hence, the Gabor transform is applied in image processing to aid detection, estimation, feature extraction, and classification.

3.2.2. Gabor Filter

A bandpass filter with a frequency response that follows the Gabor transform within a limited Gaussian interval is called a Gabor filter. Parameters such as orientation, frequency, and Standard Deviation are used to design the Gabor filter. The application of the Gabor filter for image processing has the following features:

➢

The distinct edges are highlighted using a Gabor filter.

➢

Gabor filters exhibit remarkable similarity to those observed in the human visual system, owing to parameters such as orientation, frequency, and Standard Deviation.

➢

Their ability to portray and differentiate between textures is unparalleled.

➢

Two-dimensional Gabor filters are based on a Gaussian kernel function and modulate the spatial domain with a sinusoidal plane wave.

➢

Gabor filters exhibit self-similarity as they can be produced through dilation and rotation applied to a single mother wavelet.

➢

Gabor filters can be built for various dilations and rotations and are, therefore, closely related to Gabor wavelets.

➢

Gabor filters possess appropriate spatial localization and orientation features and are localized in both the frequency and spatial domains.

The “Uncertainty Principle” governs the accuracy of the time–frequency location provided by the Gabor function. A two-dimensional circular spatial domain representation of the Gabor filter can be obtained using Equations (1) and (2) [24].

$$\left(\mathrm{x},\mathrm{y},\mathsf{\theta },\mathrm{u},\mathsf{\sigma }\right)={\displaystyle \frac{1}{2\mathsf{\pi }{\mathsf{\sigma }}^2}}\exp \left\{-{\displaystyle \frac{{\mathrm{x}}^2+{\mathrm{y}}^2}{2{\mathsf{\sigma }}^2}}\right\}\times \exp \left\{2\mathsf{\pi }\mathrm{i}\left(\mathrm{u}\mathrm{x}\mathrm{c}\mathrm{o}\mathrm{s}\mathsf{\theta }+\mathrm{u}\mathrm{y}\mathrm{s}\mathrm{i}\mathrm{n}\mathsf{\theta }\right)\right\}$$

(1)

where

$$\mathrm{x}=\mathrm{x}\mathrm{c}\mathrm{o}\mathrm{s}\mathsf{\theta }+\mathrm{y}\mathrm{s}\mathrm{i}\mathrm{n}\mathsf{\theta }$$

(2)

$$\mathrm{y}=\mathrm{y}\mathrm{c}\mathrm{o}\mathrm{s}\mathsf{\theta }-\mathrm{x}\mathrm{s}\mathrm{i}\mathrm{n}\mathsf{\theta }$$

(3)

where θ = −90° to +90° by 1°

The parameters used in the equations are explained below.

• u—sinusoidal wave frequency.
• θ—orientation angle
• σ—Gaussian function Standard Deviation

The magnitudes of the Gabor filter outputs, computed from (2), are used for subsequent analysis.

$${\mathrm{I}}_{\mathrm{w}\mathrm{\varnothing }}=\sqrt{[\mathrm{R}\mathrm{e}\left[{\mathrm{G}}_{\mathrm{w}\mathrm{\varnothing }}\left(\mathrm{z}\right)2\right]+\mathrm{I}\mathrm{m}[{\mathrm{G}}_{\mathrm{w}\mathrm{\varnothing }}\left(\mathrm{z}\right)2]}$$

(4)

Gabor filters were selected over other filters because they are better at detecting melanoma images, closely resemble the response of the mammalian visual cortex, and excel at capturing local texture, orientation, and frequency, with optimal joint localization in the space and frequency domains.

The Gabor filter is configured by a set of Control Parameters (CPs), and these CPs are used to produce the texture features from the skin dermoscopy images. The CPs of the Gabor filter are its directional orientation (ɵ), Gabor filter size of the kernel, scale or frequency (f), Gamma, bandwidth (w), and phase shift (φ). The CP is important for extracting texture features from dermoscopy images for melanoma and non-melanoma cases. The Gabor filter kernel size was set to 9 × 9 in this work after several trials to obtain optimal texture features. As ɵ represents the directional position, it is important to produce Gabor features in various directions. In this work, ɵ is set between −450 and +450. The scale (f) is set to 0.3 cycles, and the gamma parameter controls the envelope of the Gabor curve. The gamma parameter is defined as the ratio of the major axis to the minor axis of the envelope. Its bandwidth and phase shift functionally operate the Gabor filter. The Gabor bandwidth quantifies frequency selectivity and is expressed in octaves. Here, the bandwidth is set to 2.5. Filter tuning is high when the bandwidth is set to a low value, resulting in fewer feature values. Filter tuning is low when the bandwidth is set to a high value, resulting in high feature values.

4. Proposed Methodology

This work identifies and classifies melanoma disease using the Gabor-CNN (GCNN) classification approach. The dermoscopy skin image being tested is initially edge-smoothed. Subsequently, the pixels in the spatial domain are converted to a multi-resolution pixel domain via a Gabor transform. Then, the correlative matrices are computed from the Gabor-transformed image. These computed correlation-based statistical features are classified using the proposed CNN model, as shown in Figure 4a,b. The proposed CNN model uses internal and external correlation-based statistical information to classify dermoscopy images of the skin as healthy or malignant. The CNN-classified image of skin with melanoma is then segmented using a segmentation algorithm. From the segmented cancer region in the melanoma skin image, performance metrics are computed using the performance indices employed in this study.

4.1. Preprocessing

This study presents a methodology for improving the quality of source dermoscopy images to achieve more effective classification results. An enhancement algorithm increases each pixel’s grey-scale value. Numerous enhancement algorithms are available for improving skin images, such as Histogram Equalization (HE), Thresholding, and contrast enhancement. Among these image enhancement methods, HE can improve pixel regions in source dermoscopy images.

Figure 5 shows the HE-processed melanoma dermoscopy image, in which pixel intensities are enhanced by the HE method. Figure 6 shows the Gabor-transformed melanoma dermoscopy image obtained in Section 3.2.2.

Table 1. Computations of GLCM.

Parameters	Formula
Contrast	$\sum \left({\left|i-j\right|}^2\times p\left(i,j\right)\right)$
Energy	$\sum {p(i,j)}^2$
Homogeneity	${\displaystyle \frac{\sum p\left(i,j\right)}{1+|i-j|}}$
Correlation	$\sum (i-\mu i)(j-\mu j){\displaystyle \frac{p\left(i,j\right)}{[\sigma i.\sigma j]}}$
Difference Entropy	$-{\displaystyle \sum _{\mathrm{i}=0}^{\mathrm{G}-1}}{\mathrm{P}}_{\mathrm{x}+\mathrm{y}}\left(\mathrm{i}\right)\log \left({\mathrm{P}}_{\mathrm{x}+\mathrm{y}}\left(\mathrm{i}\right)\right)$
Inertia	${\displaystyle \sum _{\mathrm{i}=0}^{\mathrm{G}-1}}{\displaystyle \sum _{\mathrm{j}=0}^{\mathrm{G}-1}}{\{\mathrm{i}-\mathrm{j}\}}^2\times \mathrm{P}(\mathrm{i},\mathrm{j})$
Dissimilarity	${\displaystyle \sum _{\mathrm{i}=0}^{\mathrm{G}-1}}{\displaystyle \sum _{\mathrm{j}=0}^{\mathrm{G}-1}}\left|\mathrm{i}-\mathrm{j}\right|\mathrm{P}(\mathrm{i},\mathrm{j})$
Difference in Variance	${\displaystyle \sum _{\mathrm{i}=0}^{\mathrm{G}-1}}{\displaystyle \sum _{\mathrm{j}=0}^{\mathrm{G}-1}}{\left(\mathrm{i}-\mathsf{\mu }\right)}^2\mathrm{P}(\mathrm{i},\mathrm{j})$

presents the computation of the GLCM features used to differentiate melanoma case images from non-melanoma case images.

Table 2. Illustrations of the measured GLCM values.

Parameters	Measured Values
	Melanoma Case	Non-Melanoma Case
Contrast	$1.67\times {10}^2$	$0.061\times {10}^2$
Energy	$7.19\times {10}^3$	$1.12\times {10}^3$
Homogeneity	0.981	3.298
Correlation	0.97	0.08
Difference Entropy	0.817	4.378
Inertia	0.102	−0.861
Dissimilarity	0.761 ± 0.01	2.81 ± 0.98
Difference in Variance	0.2761	−0.9287

shows illustrations of the measured GLCM values for melanoma and non-melanoma dermoscopy images. It clearly demonstrates that the proposed melanoma detection system integrates GLCM feature extraction to distinguish melanoma from non-melanoma cases.

4.2. Feature Computations

Numerous feature-computation methods and algorithms are available for dermoscopy skin-image classification, including statistical and derivative feature sets. Traditional features are not computed from dermoscopy images across multiple directions, which degrades their value. Hence, GLCM features were used in this research to compute extensive features across multiple directions.

A statistical technique [9,14,21] for analyzing texture that accounts for the spatial relationships between pixels is computed using the grey-level co-occurrence matrix (GLCM). Its functions assess image texture by computing a GLCM. Subsequently, statistical measures are derived from the matrix, considering the rate at which pairs of pixels with a specific value and spatial relationship appear in the image. A larger GLCM indicates more greyscale levels in the original image. The GLCM values were computed from the Gabor dermoscopy image with a 90° orientation.

The expressions of the GLCM module are given in Table 1.

Here, $\mathrm{P}(\mathrm{i},\mathrm{j})$ is the Gabor dermoscopy skin image with respect to row $\mathrm{i}$ and column $\mathrm{j}$.

Table 3. Illustration of the initial setup of the hyper-parameters for the proposed CNN to classify melanoma and non-melanoma cases.

Hyper-Parameters Used for Melanoma Image Detection in the Proposed CNN	Initial Values of the Hyper-Parameters
Number of neurons in each FCL1	4096
Number of neurons in each FCL2	512
Number of neurons in each FCL3	2
Learning rate	0.0001
Drop out	0.2
Epochs	100
Decay	0.1

presents notable examples of structures for melanoma and non-melanoma images to assess the potential benefits of features for classification.

4.3. CNN Classifications

Machine learning methodologies have been used in melanoma skin cancer detection. The segmentation method required training on many images and failed to identify low-resolution images. Additionally, the machine learning algorithms required longer training on dermoscopy integumentary images. These limitations can be overcome using deep learning methods.

In the proposed CNN system, the number of Convolutional layers and GPA units was set to four, respectively. This proposed CNN comprises two modules: the Texture Feature Extractor Module (TFEM) and the Higher Feature Extractor Module (HFEM), as illustrated in Figure 5. The TFEM in the proposed CNN extracts texture features, whereas the HFEM extracts higher-dimensional features. Both the generated textures and the higher-dimensional features improve the melanoma detection rate compared with a traditional CNN. The TFEM is configured by Convolutional layers CON1 and CON2, where CON1 has 32 filters and a fixed filter size of 3 × 3, with a stride determined by the hyper-parameter selection process. Similarly, CON2 is configured with 64 filters, each with a fixed stride of 5 × 5, as determined by the hyper-parameter selection process in this work. The Gabor-transformed melanoma image is provided for CON1, and the generated two-dimensional texture features are provided for CON2 via the GPA module, which performs feature dimensionality reduction. The CON2 output values are further dimensionally reduced by GPA, yielding two-dimensional texture features. Texture features capture texture patterns in melanoma images and are important for differentiating non-melanoma from melanoma images. Figure 7: Melanoma image detection using the proposed CNN with the configurations TFEM and HFEM.

To generate two-dimensional higher-level features, the HFEM is configured with the Convolutional layers CON3 and CON4, where CON3 has 512 filters and a filter size of 5 × 5, with a stride determined by the hyper-parameter selection process. Similarly, the CON4 is configured with 512 7 × 7 filters, each with a stride, using a hyper-parameter selection process in this work. The Gabor-transformed melanoma image is fed into CON3, and the generated two-dimensional higher-order features are fed into CON4 via the GPA module, which performs feature dimensionality reduction. The CON4 output values are further dimensionally reduced via GPA, yielding higher-order two-dimensional features. Higher-order features capture the intrinsic patterns of melanoma images and are important for distinguishing non-melanoma from melanoma images. Finally, the texture features generated by TFEM and the higher-order features generated by HFEM are concatenated using the Feature Concat (FC) module, as depicted in Figure 5. It concatenates these two feature values into a single feature matrix, which is further processed by the fully connected layer (FCL). Three internal subsequence layers comprise this FCL, and each subsequence layer is configured with sets of biased neurons: 4096, 512, and 2, respectively. The biased neurons in the final layer of the FCL are used for melanoma or non-melanoma classification.

Figure 8a shows the source input melanoma dermoscopy image, and Figure 8b shows the generated feature maps from the convolutional layers of the proposed CNN.

4.4. Hyper-Parameter Optimization (HO)

The optimal selection of hyper-parameters for the proposed CNN is referred to as hyper-parameter optimization. The configuration selection is used to enhance the performance of the proposed melanoma image detection system across datasets. The researcher initially sets the hyper-parameters, and then the optimization process automatically tunes them during training. Hence, an iterative procedure is used to select the best hyper-parameters during training and testing. The number of neurons in each FCL is based on the bias weight of each neuron, decay, learning rate, dropout, epochs, number of layers, kernel size, stride position, and size of each batch. These parameters govern the entire deep learning process, optimizing functional output accuracy. The HO process employs various strategies to identify the optimal parameter values. They are reported as grid search, random search, and the Gaussian-Based Optimization Algorithm (GOA). In this research work, the GOA was used for the HO process due to its fast convergence during hyper-parameter selection of the optimal hyper-parameters. Table 3 presents the initial hyper-parameters settings for the proposed CNN to classify melanoma and non-melanoma images.

This approach helps detect melanoma by integrating the Gabor transform for textural cues with a CNN for classification. Gabor filters help identify key textural patterns in melanoma tumours, whereas CNNs classify them using complex representations. Skin imaging may more accurately identify melanoma using this hybrid strategy.

Figure 9a shows the classification output images for melanoma cancer, and Figure 9b shows the classification output images for non-melanoma cancer.

5. Result and Discussion

Table 4. Dataset specifications for melanoma skin image detection.

Dataset	Total Skin Images	Healthy Skin Images	Melanoma Skin Images
ISIC	3000	1500	1500
SIIM-ISIC Kaggle dataset	2000	1000	1000

shows the dataset specifications for melanoma skin image detection for various open-access datasets. This table shows the total number of melanoma and standard skin lesion images for each dataset.

To assess the methodology’s effectiveness, a carefully selected subset of 1500 melanoma skin cancer images and 1000 non-melanoma images was obtained from the ISIC and SIIM-ISIC Kaggle datasets. Within this subset, 750 melanoma and 250 non-melanoma skin cancer images were allocated for model training, and the remaining 750 melanoma and 250 non-melanoma images were reserved for testing. Initially curated and overseen by the University Dermatology Centre in Muncie, THIS dataset comprises 23,000 skin images spanning various melanoma cases.

The dermoscopy images in the ISIC dataset are 512 × 512 pixels, whereas those in the SIIM-ISIC Kaggle dataset are 1024 × 1024 pixels.

This extensive collection is valuable for training and validating melanoma detection algorithms. An enhancement method is required to distinguish skin–hair pixel values from those in the region of the malignancy boundary. Figure 10a is a dermoscopy image of the skin, and Figure 10b is an improved version of the same image using the Histogram Equalization (HE) technique.

These results represent the outcomes of the recommended melanoma detection method, which was successfully applied and evaluated on the publicly available skin image datasets ISIC and SIIM-ISIC. The results, including accuracy, specificity, and sensitivity, are presented and analyzed, demonstrating the superiority of the proposed approach over existing methods. The graphical user interface (GUI) was developed in MATLAB R2020 on a computer with 16 GB of RAM and an Intel Core i5 processor.

The efficiency of the proposed melanoma detection system was evaluated using several important parameters, as shown in Equations (5)–(7). The objective of these parameters is to evaluate the system’s efficacy and precision in detecting melanoma.

Sensitivity (Se) = TP/(TP + FN)

(5)

Specificity (Sp) = TN/(TN + FP)

(6)

Accuracy (Acc) = (TP + TN)/(TP + FN + TN + FP)

(7)

where

• “TP: True Positive—Correctly identified cancer pixels.”
• “TN: True Negative—Correctly identified non—cancer pixels.”
• “FP: False Positive—Wrongly identified cancer pixels.”

Figure 8 shows the specifications of correctly detected melanoma and healthy skin images for various datasets. In this study, the proposed system correctly detected 1497 of 1500 healthy skin images and 1495 of 1500 melanoma skin images in the ISIC dataset, as shown in Figure 11.

Figure 12a shows the confusion matrix for the ISIC dataset, and Figure 12b shows the confusion matrix for the SIIM-ISIC dataset.

Table 5. Comparisons of the proposed method with other methods using quantitative metrics.

References	SIIM-ISIC Kaggle			ISIC
	Sensitivity (Se)	Specificity (Sp)	Accuracy (Acc)	Sensitivity (Se)	Specificity (Sp)	Accuracy (Acc)
Proposed work	98.63	99.11	98.77	98.66	98.75	98.56
Arvind Kumar Shukla et al. (2025) [25]	97.12	97.98	97.27	97.65	97.16	97.64
Abohashish et al. (2025) [26]	97.10	97.05	97.34	97.61	96.15	96.58
Al-Rasheed et al. (2023) [11]	96.38	96.67	96.75	95.83	95.68	95.74
Khan et al. (2019) [17]	95.57	95.87	95.86	95.37	95.28	95.85
Yu et al. (2017) [12]	94.67	94.28	94.85	94.39	94.76	94.27

presents the performance of the proposed melanoma detection system on the SIIM-ISIC Kaggle and ISIC datasets. The proposed system achieves 98.63% and 98.66% Se, 99.11% and 98.66% Sp, and 98.77% and 98.58% Acc on the ISIC and SIIM-ISIC Kaggle datasets, respectively. The proposed Gabor feature-based NN model presented in this work achieves higher cancer segmentation results due to its feature extraction and computation from images. Gabor computed features, including GLCM, improve cancer segmentation accuracy by providing pixel stability relative to surrounding pixels, thereby increasing segmentation accuracy within the cancer region. To validate the robustness of the proposed model, it was applied to various dermoscopy image datasets, and the results are presented in

Table 6. Experimental results in the dermoscopy imaging dataset.

Images	Experimental Results in % SIIM-ISIC Kaggle Dataset			Experimental Results in % ISIC Dataset
	Se	Sp	Acc	Se	Sp	Acc
1	98.8	99.3	98.4	96.9	97.9	98.5
2	98.5	99.4	98.8	97.8	98.9	98.3
3	98.7	99.3	99.3	98.6	98.6	97.9
4	99.3	98.7	99.4	98.6	99.3	97.6
5	98.7	98.9	98.7	98.9	98.7	98.3
6	98.4	99.3	98.4	98.9	98.5	98.7
7	98.3	99.2	98.8	98.5	98.7	98.6
8	98.8	99.7	98.3	99.4	99.3	99.4
9	98.5	98.7	98.2	99.3	98.2	99.3
10	98.3	98.6	99.4	99.7	99.4	99.2
Average	98.63	99.11	98.77	98.66	98.75	98.58

.

Arvind Kumar Shukla et al. (2025) [25] used a banyan tree growth Optimization Algorithm for melanoma image detection, leveraging a significant set of texture features. Abohashish et al. (2025) [26] used an LSTM-based classification algorithm to differentiate melanoma images. Al-Rasheed et al. (2023) [11] used a non-linear modelling algorithm to derive a systematic set of features for classification. Khan et al. (2019) [17] enhanced the contrast of skin lesions for saliency mapping and the Particle Swarm Optimization (PSO) algorithm to identify accurate borders. Yu et al. (2017) [12] modelled a skin lesion detection system using a CNN.

The importance of the Gabor transform and the proposed CNN is illustrated in

Table 7. Comparisons of the importance of Gabor and the proposed CNN for the melanoma image detection system on the SIIM-ISIC Kaggle dataset.

CNN Methods for Melanoma Image Detection Process	Measured Values in %
	Se	Sp	Acc
Gabor + Proposed CNN (in this paper)	98.63	99.11	98.77
Proposed CNN only	96.12	97.87	96.14
Gabor + traditional CNN	95.65	95.28	95.71
Contourlet + Proposed CNN	94.87	94.15	94.28
Curvelet + Proposed CNN	93.18	93.85	93.16

and

Table 8. Comparisons of the importance of Gabor and the proposed CNN for the melanoma image detection system on the ISIC dataset.

CNN Methods for Melanoma Image Detection Process	Measured Values in %
	Se	Sp	Acc
Gabor + Proposed CNN (in this paper)	98.66	98.75	98.58
Proposed CNN only	96.27	96.76	96.16
Gabor + traditional CNN	95.98	95.18	95.64
Contourlet + Proposed CNN	95.16	95.01	95.12
Curvelet + Proposed CNN	94.09	94.28	94.01

. In this work, a traditional CNN and a proposed CNN were combined with Gabor, Contourlet, and Curvelet filters, and the experimental results are reported. Table 6 compares the importance of the Gabor filter and the proposed CNN for the melanoma image detection system on the SIIM-ISIC Kaggle dataset.

Table 8 compares the importance of the Gabor filter and the proposed CNN for the melanoma image detection system on the ISIC dataset.

Based on detailed experimental analyses of the proposed method and similar recent dermoscopy cancer-detection methods, this work achieves higher performance due to its stability across a range of dermoscopy images and pixel resolutions. Conventional cancer detection in dermoscopy images relies on pixel intensities, leading to suboptimal cancer segmentation for low-intensity pixels. To eliminate such limitations in conventional dermoscopy-based cancer detection methods, this work proposes a Gabor-based CNN model. This proposed model is evaluated on both low- and high-intensity dermoscopy images, demonstrating the algorithm’s stability across multiple performance metrics.

Table 9. Comparisons of the proposed method with other methods using qualitative metrics.

Qualitative Metrics	Proposed Work	Arvind Kumar Shukla et al. (2025) [25]	Abohashish et al. (2025) [26]	Al-Rasheed et al. (2023) [11]	Khan et al. (2019) [17]
Method	Proposed CNN	Ensemble classifier	Traditional CNN	AlexNet CNN	SVM classifier
Data source	ISIC and SIIM-ISIC	ISIC	Kaggle dataset	ISIC	Kaggle dataset
Training time	low	moderate	High	moderate	Moderate to high
Interpretability	low	moderate	moderate	moderate	High
Robustness	High	moderate	moderate	low	Low

compares the proposed method with other methods across qualitative metrics, methods, datasets, training time period, interpretability, and robustness.

Error rate estimation is the process of analyzing the performance of the proposed melanoma detection system. The error rate is inversely proportional to the accuracy rate and is expressed as a percentage. In this paper, the error rates for the ISIC and SIIM-ISIC datasets are 1.23% and 1.44%, respectively. Based on the analysis of error rates across datasets, the proposed method exhibits higher accuracy and lower error rates.

6. Conclusions

A Gabor filter-based CNN for skin cancer classification is proposed in this paper. The proposed system uses two categories of open datasets from SIIM-ISIC and ISIC to detect and delineate tumour regions in the skin of individuals with melanoma, employing a computer-aided system with a graphical user interface. The present study employed Histogram Equalization for contrast enhancement and decomposed skin images into transformed coefficients. Furthermore, histogram coefficients facilitated the extraction of various characteristics, including contrast, homogeneity, energy, correlation, difference entropy, inertia, dissimilarity, and variance. The attributes above were trained and classified using a deep learning Gabor-transform-based classifier to determine whether the captured images correspond to melanoma cases. The results of this study demonstrated a sensitivity (Se) of 98.64%, specificity (Sp) of 98.93%, and an overall accuracy of 98.67%. The results were compared with existing work and found to outperform those of Arvind Kumar Shukla et al. (2025) [25], Abohashish et al. (2025) [26], Al-Rasheed et al. (2023) [11], Khan et al. (2019) [17], and Yu et al. (2017) [12]. Subsequent research endeavours will focus on applying the enhanced model of the proposed methodology to address lung cancer and achieve optimal outcomes.

The limitations of this study are as follows.

• The proposed system has not been verified using clinical dermoscopy images; hence, its robustness is not validated in this work, which is identified as the main drawback of this research.
• In this study, only melanoma images are detected, including their affected regions, but severity levels were not determined.

In the future, the performance of the proposed melanoma image detection system can be enhanced by implementing transformer-based classification algorithms, thereby yielding significant improvements in results. Moreover, the clinical dermoscopy images will be evaluated by the proposed system described in this article to assess its robustness in future studies.