Artificial Intelligence based real-time automatic detection and classification of skin lesion in dermoscopic samples using DenseNet-169 architecture

Abstract

The most common challenge faced by dermoscopy images is the automatic detection of lesion features. All the existing solutions focus on complex algorithms to provide accurate detections. In this research work, proposed Online Tigerclaw Fuzzy Region Segmentation with Deep Learning Classification model, an intellectual model is proposed that provides discrimination of features with classification even in fine-grained samples. This model works on four different stages, which include the Boosted Anisotropic Diffusion filter with Recursive Pixel Histogram Equalization (BADF-RPHE) in the preprocessing stage. The next step is the proposed Online Tigerclaw Fuzzy Region Segmentation (OTFRS) algorithm for lesion area segmentation of dermoscopic images, which can achieve 98.9% and 97.4% accuracy for benign and malignant lesions, respectively. In the proposed OTFRS, an accuracy improvement of 1.4% is achieved when compared with previous methods. Finally, the increased robustness of lesion classification is achieved using Deep Learning Classification –DenseNet 169 with 500 images. The proposed approach was evaluated with accuracy classifications of 100% and 98.86% for benign and malignant lesions, respectively, and a processing time of less than 18 sec. In the proposed DensetNet-169 classification technique, an accuracy improvement of 3% is achieved when compared with other state-of-art methods. A higher range of true positive values is obtained for the Region of Convergence (ROC) curve, which indicates that the proposed work ensures better performance in clinical diagnosis for accurate feature visualization analysis. The methodology has been validated to prove its effectiveness and throw light on the lives of affected patients so they can resume normalcy and live long. The research work was tested in real-time clinical samples, which delivered promising and encouraging results in skin cell detection procedures.

Keywords

Boosted Anisotropic Diffusion filter with Recursive Pixel Histogram Equalization (BADF-RPHE)Deep learning Classification - DenseNet 169 Proposed Online Tigerclaw fuzzy Region Segmentation (OTFRS)Skin tumor

1 Introduction

Melanoma acts as the anomalous and deadliest cancer disease, which is caused by uncontrolled growth in skin cell structure. The skin color is produced by peculiar cells called melanocytes for melanin pigment production. Benign lesions are curable when detected at an early growth stage. The main cause of skin lesions is ultraviolet rays, which act as a carcinogen in the human body. The report provided by the American Cancer Society indicates that 89,070 cancer cases will be diagnosed in the United States in 2022 [1]. The non-melanoma lesions hold about 96% of all skin cancers, including squamous and basal carcinoma types. According to the International Cancer Institute, the 5-year survival rate is 26.3% [2].

Initially, the time required for skin lesion pattern description was very tedious. Most often, the cancer cells are also misinterpreted as healthy cells during lesion removal. This causes the unhealthy cells to grow abruptly, forming new skin lesions. It spreads to entire body regions, which affects the overall functioning of the body. Hence, a highly accurate and efficient automatic system is needed for lesion segmentation and classification. Since the skin lesions are similar in appearance, it is difficult to classify the lesions. A gold standard for radiologists to access these lesions is dermoscopic samples. For such effective analysis and diagnosis, there are many ongoing active research projects on these skin lesions. Dermatologists and oncologists can see the effectiveness of the proposed research during surgery on these skin lesion cells.

OTFRS is the novel proposed segmentation method used in our research work, which effectively contributes to the effective bifurcation of lesion regions in the skin sample. This is similar to the use of tiger claws for catching prey. Thus, continuing hand in hand with the previous statement, the proposed new algorithm segments the lesion cells, which follow the patching of similar pixels along with fuzziness to retain the information without being lost for improved effective segmentation. The classification is done by the deep learning DenseNet-169 structures, which train the feature vectors based on lesion regions. Hence, the radiologists must detect the cell lesion followed by classification, which has to be performed with greater care. A range of reliable techniques finds effectiveness in the screening of affected cells. Effective implementation of this research work is done concerning such standard journals and studies. The shape of the skin lesion varies due to greater heterogeneity and variation. Taking this feature as a key aspect, the proposed work helps in the detection of tumors even in the micro environment.

2 Related works

The earlier method of lesion classification suffered from insufficient data and processing tools. At earlier stages, only manual lesion extraction was available, which resulted in false diagnoses and increased the time scale range. The advancement of automatic lesion extraction systems eradicated such examination problems.

Harangi et al. developed a classification based on weighted convolutional systems [3] with the transfer learning scheme. The accuracy for cancerous segmentation is 91.3% and for classification is about 89.1%. This creates a fusion-based classification of lesion cells. Fikret et al. pointed out Artificial Neural Network (ANN) system classification [4] with 80% accuracy with novel feature discrimination. Thomas et al. developed semantic segmentation with an identification rate of 92.5% for classification [5], which is entirely based on auto-generated spatial geometric analysis. This method has an accuracy range of 93.6 to 97.9% for lesion cell classification. The back-propagating conjugate resilient algorithm [6] was created by Masood et al. This method has a specificity and sensitivity of 95.1% and 92.6%, respectively.

Harika et al. developed a texture distribution method [7] that accurately classifies lesion cells from healthy cells with a segmentation accuracy of 95%. The identification rate of 94% was with pipeline deep learning architecture [8]. This method was proposed by Masood et al. and developed U-Net with semantic segmentation for lesion segmentation of skin cells. Afshar et al. developed a capsule classification network with an accuracy of 90.89%. This is accompanied by gabor filter detection [9] for preprocessing of input image features. A sensitivity range of 96% was obtained with image reconstruction techniques. This technique was proposed by Satheesha et al. which helps for effective diagnosis [10].

An optimization approach called moth flame [11] was developed by Khan et al. with deep saliency segmentation of 95.38% and a classification kernel classifier of 90.67% to determine variant class feature labels. Seiffert et al. provided a variance-based rusboost classification system [12] with spectral properties. It is also a sample randomized ada-boost system with an accuracy of 94.12%. 80.14% accuracy was obtained with topological multiple image pixel representation by Luminata et al. This provides the level classifier set [13]. Singh and Mary developed lesion classification based on the support vector machine concept to feature vectors effectively with an accuracy of 93.18% [14]. ISIC 2017, PH2, and the HAMN 1000 dataset are used in all these related works.

2.1 Research gap recognized

The complexity of other systems increases with an increase in accuracy [15, 16]. The inaccurate classification of these tumor cells also accounts for the premier reason for the increase in the mortality rate. The error-based values that occur due to an increase in noise sensitivity values are also the main reason for misclassifications [17]. The validation of these models is very difficult to establish after every set of operations as the frequency range entirely depends on the histogram of images [18]. The overlapping of lesion cells causes sensitivity to scalability in training the data sequence [19]. It also lacks in combining the disparate system feature representations, ignoring the spatial-spectral relationships among the patches in the class [20].

2.2 Contributions to the research work

A Boosted Anisotropic Diffusion Filter indicates a reduced time on edge preservation. This reduces vision loss, whereas the overall surveyed filtering technique does not, as in [21].

Recursive Pixel Histogram Equalization provides the technique of equalization, which does not provide intensive computation, whereas existing research shows indiscriminate values, which increase noise at the background level [22].

The novel proposed system of Online Tigerclaw Fuzzy Region Segmentation enables efficient spatial consistent features by finding unique patches automatically, following the case in which the tiger uses the claws for prey catching. This eliminates boundary overlap as in [23].

DenseNet-169 based Deep Learning Classification is performed, which proves to be a novel procedure for lesion classification. This method increases classification accuracy with higher true positive values on the ROC curve, as described in [23 –25].

An in-depth comparative analysis of performance is done on segmentation and classification methods [33 –35] with existing novel research techniques. These standard measures prove the proposed system’s effectiveness. This novel work can accurately and effectively aid radiologists in analyzing the skin lesion at an early stage.

The organization of the paper: In Section 3, the proposed methodology is discussed. Performance evaluation and results are under Sections 4 and 5. Comparison of classifiers in Section 6. Section 7 concludes the paper.

3 Proposed methodology

The novel proposed segmentation technique OTFRS, followed by deep learning classification using DenseNet-169, is proposed. The technique of automatic extraction and classification of lesion cells is given in Fig. 1. Stage (i) (Acquisition): The input image is obtained from the ISIC dataset. Stage (ii) (preprocessing): The acquired image is subjected to BADF, followed by Hue Saturation Value (HSV) color change, and then RPHE. Stage (iii) (Segmentation): The lesion regions are segmented using the proposed OTFRS from the non-lesion regions. This stage is accompanied by Contour Superpixel Segmentation (CSS) to determine the super intensity of pixels in the image. Stage (iv) (Classification): Deep Learning Classification with DenseNet-169, which provides creditable output based on the Asymmetry-Border-Color-Diameter (ABCD), and Gray Level Co-occurrence Level (GLCM) features. The ROI curve is extracted from a skin lesion for the classification of nevi or melanoma. The same range of sequential streams is adopted for test and training sets of data in the proposed novel research work.

Fig. 1

Proposed research work flow of Skin lesion Classification Algorithm.

3.1 Stage (i): Image Acquisition

The research work relies on the ISIC archive dataset [26], which holds 450 benign and 450 malignant skin lesions. This benchmark image dataset has PNG, JPEG, JPG, and DICOM formats. In addition, the research work is performed even on clinical images —95 patients based on 520 computerized examinations —from International Cancer Centre (ICC), Neyyoor, Tamil Nadu. It has 36 of 95 high-risk skin samples for lesion detection and classification.

3.2 Stage (ii): Preprocessing of the acquired image

The premier goal of this particular stage is to enable the enhancement of lesions in various categories. Anisotropic notch diffusion filter effectively eliminates noise for better and more valuable interpretation. The average range of pixel values is given in Equation (1) as follows, $A (p) = \sqrt{\sum_{p = 1}^{q} {(p * \bar{P})}^{2} / t - 1}$ (1) where, p is the value of the assigned pixel, $\bar{P}$ is the average pixel value, and t is the total pixel. The algorithm for AND-RPHE is given as,

Algorithm 1: Enhancement through Boosted Anisotropic Diffusion Filtering-Recursive Pixel Histogram Equalization (BADF-RPHE) Algorithm
Step 1: Initially, the input image is subjected to the transformation of color intensity values with multiple resolution representations, which are split into four individual bands. The intensity (p,m) of the transformative vector is given as,
$I_{t} = Tr (I (p, m)) (2)$
Step 2: Mapping with color functions accounts for effective color picking at individual, unique clusters.
Step 3: Based on the cluster’s common value, a seed point is selected from the lesion points.
Step 4: Initial binary-mapped multiplication is done to perform the filling of holes in the lesion part.
Step 5: Recursive Pixel Histogram Equalization increases the contrast level of these pixels individually.
Step 6: Soft thresholding sets the value to zero when it is less than the threshold value, or it moves to the next value and is tested. The process is repeated until all sub-bands are completed.

The low-value contrast in the acquired image creates unclear visualization [27] during the research work. In this research work, Boosted Anisotropic Diffusion Filtering for speckle noise elimination is used. BADF-RPHE provides visual enhancement that shows its effectiveness even with variations in illumination, shade, and intensity color values. During medical diagnosis, the RGB-to-HSV color change plays a prominent role. The false rate of illumination is reduced in the chrominance and luminance sections.

3.3 Stage (iii): Proposed OTFRS-CSS algorithm for segmentation

The main feature of the OTFRS-CSS segmentation algorithm is to increase its efficiency through the effective removal of foreground objects from their backgrounds. This novel method of segmentation is intended to increase accuracy while reducing processing time [29]. The node regions are calculated based on the centroid of the candidate that is located. The OTFRS algorithm identifies the tumors from the lesion images using a combination of online tigerclaw region based segmentation and a fuzzy region segmentation algorithm. OTFRS with the CSS algorithm shows unique and interesting results for diagnosis by dermatologists in the medical field. Superpixels contain more information than pixels and align with image borders better than rectangular image patches. These help in the grouping of pixels. The superpixel representation significantly decreases the number of image primitives when compared to the conventional pixel representation in an image, increasing representational efficiency. This research work suggests an enhanced segmentation algorithm for superpixel grid computed tomography (CT) images employing active contours using the idea of superpixel segmentation.

The determination of the initial contour is more precise when super pixels are used to enhance the region growth criterion. The paper boosts the precision of skin lesion segmentation and enhances image quality by fully taking into account the pixel texture features. Then, using fuzzy based region mapping techniques for image segmentation, regions are analyzed based on region and edge information in the picture. In a manner similar to how tigers use their claws to capture prey when on the hunt for food, the technique of Online Tigerclaw Fuzzy Region Segmentation (OTFRS) is used to separate tumor cells from healthy cells, eliminating overlapping borders.

Algorithm 2: Proposed Online Tigerclaw Fuzzy Region Segmentation- Contour Superpixel Segmentation (OTFRS-CSS) Algorithm
Step 1: Image vertices are created using the pixels and the neighborhood pixels, which denote the edges of the image. This is similar to the graph with the edges and vertices, respectively. The image vertices using distance measurement are given by,
$E (k 1, k 2) = {((k 1_{2} - k 1_{1})^{2} + (k 2_{2} - k 2_{1})^{2})}^{1 / 2} (3)$
Step 2: Superpixel values are chosen based on the priority values of the lesion images. Two seed points are chosen, one of which indicates skin, and the other with lesion regions is highlighted [29].
Step 3: A probability value is chosen based on the fixed pixel value of the image. The superpixel values are chosen by weight-based functional values, which are given by,
$w (k, M_{i}^{D}) = exp (- (\frac{d (M (k), M_{i}^{D} (k))}{U^{2}} + M (k)) (4)$
Where, d(.,.) is the distance on weighted euclidean measures and U² is the maximal distance on selected seed pixels in the dataset.
Step 4: Based on the selected superpixel lesion values, the global patch region for the pixel region is selected. Each mapped label has a higher membership probability, where label 1 indicates a lesion and label 0 if the pixel values indicates skin-based normal regions. The decision probability for a lesion is described by the following Equation (18),
$U (sl \| k) = \frac{\sum_{k 2 \in U_{k}} \sum_{i = 1}^{n} w (k 2, M_{i}^{D} (k 2)) . k_{i}^{D} (k 2, n)}{\sum_{k 2 \in M_{x}} \sum_{i = 1}^{n} w (k 2, M_{i}^{D} (k 2))} (5)$
Where, $M_{i}^{D} (k 2, n)$ is a labelled patch value assigned to x.
Step 5: From the mapped region pixels, the prior pixel values are segmented using the tiger claw concept by effective removal of lesion regions similar to the claws in a tiger which tears the prey’s skin.
Step 6: Effective segmentation is done by extracting the normal region from lesion regions by pixel value labelling lesions on dataset images.

3.4 Stage (iv): Deep Learning Classification –DenseNet 169 (DLC-DenseNet 169)

For the effective classification of skin lesion images, DLC- DenseNet 169 is used. This method uses data based on a labelling scheme with various levels of abstraction. Hence, it is also called the AI-DLC technique. This DenseNet-169 DLC enables faster and easier training sequences that automatically train the data with feature values on several images. The novelty of using the proposed DenseNet 169 is that it requires far fewer training parameters, which reduces the processing time. For the skin lesion diagnosis, feature values are automatically extracted. These extracted feature values depend on the GLCM, ABCD, and FOS vector values. The neural classification network is employed with layers corresponding to input, hidden, and output layers. The proposed system of lesions on dermoscopy images is brought forward by the DenseNet-169 DLC structure, with 5 layers. The dimension is getting bigger at every layer because the feature maps required for processing are concatenated eventually. The growth rate, termed as hyper-parameter, controls how much data is added to the network at each layer with respect to obtained skin tumour dataset. These parameters are established prior to the training. Moreover, the system’s ability to learn even more complicated representations is strongly impacted by the higher number of layers of the classifier systems. The steps for tuning is given below:

Establish a search area: During the tuning process, the range of values are determined for individual hyper-parameter.

Choose a tuning technique: Select a method to investigate the search space and assess how well various hyper-parameter settings perform.

Create evaluation standards: Establish a statistic or metrics to judge how well each model configuration performs. Precision, Accuracy, Sensitivity, Specificity validation loss are typical measurements.

Hyper-parameter search: It is carried out by training and comparing numerous models with various combinations of hyper-parameters. Here, each model is trained on a training set, its performance is assessed on a validation set, and the best-performing configuration is chosen.

Evaluation of a held-out test set performance: It is essential to assess the final model’s performance on a different skin tumour set after the best hyper-parameter configuration has been identified. This offers a neutral assessment of the model’s capacity for generalization.

The classification is entirely based on the testing and training phases, where its effectiveness is determined by the following proportions: 90% -10%, 80% -20%, 70% -30%, and vice versa. The grouping of the dataset is to prove the effectiveness of classifiers, with the best case at 90% -10% and the worst case at 10% –90%, respectively.

(i) Class Feature Extraction -GLCM

Class feature extraction on the gray-level co-occurrence matrix is done automatically by the DLC classifier [30], depending on the grey and spatial values. This research work enables the following features: energy, contrast, correlation, and homogeneity. $Energy / E = \sum_{f = 0}^{s - 1} \sum_{p = 0}^{s - 1} x^{2} (f, p)$ (2) $Contrast / C = \sum_{f = 0}^{s - 1} \sum_{p = 0}^{s - 1} {(f - μ)}^{2} x (f, p)$ (3) $Correlation / Cor = \sum_{f = 0}^{s - 1} \sum_{p = 0}^{s - 1} \frac{(f - μ) (p - μ) x (f, p)}{σ_{f} σ_{p}}$ (4) $Homogeneity / H = \sum_{f = 0}^{s - 1} \sum_{p = 0}^{s - 1} \frac{x (f, p)}{1 + | f - p |}$ (5)

(ii) Class Feature Extraction –GLCM

The ABCDs of dermoscopy images indicate asymmetry values, border, color, and also the diameter of the tumor. Equal patterns of skin tumors are termed as symmetric patterns. By using the degree pattern, the asymmetric values can be effectively calculated.

Asymmetric value: The value of AV is calculated by dividing the major and minor areas with Equation (10), $Av = ({Av}_{1} + {Av}_{2}) / 2 {Av}_{total}$ (6)

Where, Av₁ and Av₂ represent minor and major areas of a segmented, non-overlapping region. Av_total is the total tumor area implemented.

Border: Edge sharpening, compact values, fractal indexing, and gray transition are various metrics to calculate border irregularity values.

Edge sharpening is based on maximum axis and minimum axis length.

$ES = ((MaxMin) 5 % + 2) / 100$ (7)

A compact range is indicated based on lesion area (La) and lesion perimeter (Lp).

$CR = 2 Lp = 4 La$ (8)

Fractal Indexing is implemented using box fractal values. The box counting technique represents block numbers, with the hausdroff plot indicated by the logarithm of fractal values and block numbers.

The gray transition is formulated based on diameter and direction. The gradient operation can be indicated using the border values. These transitions are determined by using the patch-based masking scheme. The mean and variance are also calculated.

Color value: The color of dermoscopy patches is dependent on the intensity of the tumor intensity sample implemented for research purposes. The effect on tumors depends on the larger pixel intensity, and vice versa.

Diameter: The axis length with the minimal value is used to determine the diameter of the region function. The malignant regions indicate non-uniform regions for further processing.

Algorithm 3: Testing: DLC –DenseNet 169
Input state vector 1: DLC Classifier - DenseNet 169 //*5 level set*//
Input state vector 2: Threshold //***3 elemental array for trained features i.e., GLCMIndex, FOSIndex, CFIndex****//
Input state vector 3: Sample //**Vector testing**//
Output state: Labelvalue //**Boolean range-True or False Positive **//
For l = 1 to 3
Begin
[Fs0,Fs1] = phase (test) [classifier (l), Sampling value] //** prediction score**//
If fn<Threshold value(l) then Begin
Outstate label range = False Positive; Exit;
End //**** If loop ends ****//
End //**** For loop ends *****//
[Fs0,Fs1] = phase (test) [DLC classifier (l), sampling value];
Outstate label = (Fs1 > Fs0)

Algorithm 4: Training: DLC –DenseNet 169
Input*vector 1 - Fst = Feature set training
Inputvector 2 - GLCM_indexrange = GLCM features - extraction in Fst
Inputvector 3 - MF_indexrange = Main features - extraction in Fst
Inputvector 4 - FOS_indexrange = FOS features - extraction in Fst
Inputvector 5 - SF_indexrange = Structural features - extraction in Fst
Output state label_1: DLC Classifier- DenseNet 169 //***5 level set***//
Output state label_2: Range of threshold //* 4 features i.e., SF _ values, GLCM _ values, FOS _ values, MF _ values *//
Index based value = [GLCM_ indexrange, MF_ indexrange, FOS_ indexrange, SF_ indexrange]
L = Total training cases
T = 2; //Number of classification by DLC for images// Threshold = [0 0]; //Threshold initialization//
For g = 1 to T
Begin
[Fs0, Fs1, DLC Classifier (g)] = predict (“State of DL”, Fst Index value (g)]; // **** Training feature set (Tfs)-*
Training and evaluation with index (g)***//
Normalize [Fs0,Fs1];
Td = 0;
While (Td < = 1)
Begin
Outstate label = (f1> = Td) //**Binary label**//
If(Label = false_negative)
Begin
V = Td-0.01;
Break the loop;
End//*** If loop ends**//
Else
V = Td+0.01;
End//***End: while loop***// Threshold (g) = Td;
Update value (Tls = Classifier (g), Threshold (g); //** false positive: Tls and repeated state Tls **//
End//**For loop ends**//
[Fs0, Fs1, Classifier (g)] = predict (“DL”, Tls Index value (g)];

(iii) Extraction of statistical features –FOS

Color features depend on mean, kurtosis, standard deviation, entropy, and skewness. The colors indicate the FOS features in the image information content. $Mean = μ_{n} = 1 / Z \sum_{x = 0}^{z - 1} f_{nx}$ (9)

$Kurtosis = λ_{n} = {[1 / Z \sum_{x = 0}^{z - 1} {(f_{nx} - μ_{f})}^{4}]}^{1 / 4}$ (10)

$Std . Deviation = σ_{f} = {[1 / Z \sum_{x = 0}^{z - 1} {(f_{nx} - μ_{f})}^{2}]}^{1 / 2}$ (11) $Entropy = E_{f} = - \sum_{x = 0}^{z - 1} {(f_{x} * log f_{xf})}^{2}$ (12) $Skewness = θ_{f} = {[1 / Z \sum_{x = 0}^{z - 1} {(f_{x} - μ_{f})}^{3}]}^{1 / 3}$ (13)

Where, Z is the total range of pixels with the color intensity f_x of the x^th pixel. The value zero corresponds to benign or non-tumor tissue, while one is classified as a tumor or malignant.

Samples for testing

When a false positive case is determined by the classifier, then the label points to the same value. This exhibits a true state when the testing of the classifier corresponds to a true state. The structural and corresponding feature values are classified correctly, even in the overlapped lesion regions.

Samples for testing

The class sample determination with a threshold value is the main goal for the sample to be trained in the database. Two scores of predictions are made by the classifier. The higher the class value, the higher the chance of a prediction. Fs0 and Fs1 point to false positive and true positive class scores, respectively. The 0 to 1 range of values corresponds to GLCM, structural, and FOS class features. The main features are based on the discriminatory properties of mean, correlation, standard deviation, kurtosis, skewness, variance, etc.

4 Statistical performance analysis

The standard performance metrics of histogram values relate to the value of mean brightness error (MBE) values based on the mean of Input (Z_n) and output (X_n) of the image values and PSNR values depend on levels (P), and Mean Square Error (MSE). $Mean Brightness Error = (Z_{n} - X_{n})$ (14) $PSNR = 10 {log}_{10} (P - 1)^{2} / MSE$ (15)

The mean square error and correlation energy serve as the performance metrics for the segmentation algorithm for the samples taken for processing. The mean square error is calculated by the following equations, $Mean Square Error = \sum_{n 1 c 1} \frac{[I (n 1, c 1) - I (n 2, c 2)]}{P 1 \times N 1}$ (16) $Root Mean Square Error = \sqrt{MSE}$ (17) $Correlation Energy = \frac{\sum_{z} \sum_{x} [N_{zx} R_{d} [z, x]] - μ_{z}^{2}}{σ_{z} σ_{x}} c$ (18)

Where, μ_z = ∑R_d (z, x) and $σ_{z}^{2} = \sum z^{2} R_{d} [z, x] - μ_{z}^{2}$

The overall energy of the pixels in tumor cells is specified by the equation (23),

$L_{e} = \sum_{1}^{N} \sum_{1}^{L} | N (z, x) |$ (19)

5 Results and discussion

The research work was carried out on a Windows PC; Platform: MATLAB 2018a; RAM capacity: 4 GB; CPU speed: 2.6 GHz. The Boosted Anisotropic Diffusion Filter - Recursive Pixel Histogram Equalization- (BADF-RPHE) method provides the effectiveness of the proposed research work. The super pixel value enhancement is done to improve the contrast values. The CSS concept effectively extracts the frequency value at higher values. Various equalization methods with different performance metrics for various benign and malignant lesion images are compared in Table 1.

Table 1
Comparison of preprocessing-histogram equalization for various images

Stage Sample Images DSIHE BBHE MMBEBHE CSRSHE RMSHE RPHE

Absolute Mean Benign 20 15.24 15.22 14.42 6.85 3.35 3.25

Brightness Error 133 25.51 4.64 3.77 5.43 1.27 1.24

270 12.78 15.44 9.17 4.76 2.17 2.19

Malignant 353 12.88 27.34 37.78 19.73 0.14 0.11

381 45.22 46.65 38.85 14.25 4.31 4.27

427 57.61 25.37 27.75 15.75 3.81 3.79

Average 28.21 22.44 21.96 11.07 2.51 2.47

Standard Benign 20 72.48 52.44 63.81 54.79 42.55 42.49

Deviation 133 72.65 58.25 77.84 56.52 41.84 41.83

270 78.47 68.41 63.32 63.94 31.57 31.55

Malignant 353 81.96 73.55 77.76 62.74 34.15 32.56

381 84.22 66.24 81.14 65.27 36.77 37.66

427 81.14 72.88 83.86 57.23 29.66 28.77

Average 78.48 65.29 74.62 60.08 36.09 35.81

Peak Signal To Benign 20 14.16 15.38 22.25 27.36 31.87 33.87

Noise ratio 133 12.28 15.37 17.54 28.55 35.82 35.88

270 16.54 11.66 12.37 20.77 39.39 42.29

Malignant 353 18.57 12.76 14.44 21.46 37.55 38.44

381 16.55 19.55 21.84 23.26 37.88 35.98

427 15.83 17.35 21.75 24.39 38.84 38.95

Average 15.66 15.35 18.37 24.29 36.89 37.57

Stage	Sample Images	DSIHE	BBHE	MMBEBHE	CSRSHE	RMSHE	RPHE
Absolute Mean	Benign	20	15.24	15.22	14.42	6.85	3.35	3.25
Brightness Error		133	25.51	4.64	3.77	5.43	1.27	1.24
		270	12.78	15.44	9.17	4.76	2.17	2.19
	Malignant	353	12.88	27.34	37.78	19.73	0.14	0.11
		381	45.22	46.65	38.85	14.25	4.31	4.27
		427	57.61	25.37	27.75	15.75	3.81	3.79
	Average		28.21	22.44	21.96	11.07	2.51	2.47
Standard	Benign	20	72.48	52.44	63.81	54.79	42.55	42.49
Deviation		133	72.65	58.25	77.84	56.52	41.84	41.83
		270	78.47	68.41	63.32	63.94	31.57	31.55
	Malignant	353	81.96	73.55	77.76	62.74	34.15	32.56
		381	84.22	66.24	81.14	65.27	36.77	37.66
		427	81.14	72.88	83.86	57.23	29.66	28.77
	Average		78.48	65.29	74.62	60.08	36.09	35.81
Peak Signal To	Benign	20	14.16	15.38	22.25	27.36	31.87	33.87
Noise ratio		133	12.28	15.37	17.54	28.55	35.82	35.88
		270	16.54	11.66	12.37	20.77	39.39	42.29
	Malignant	353	18.57	12.76	14.44	21.46	37.55	38.44
		381	16.55	19.55	21.84	23.26	37.88	35.98
		427	15.83	17.35	21.75	24.39	38.84	38.95
	Average		15.66	15.35	18.37	24.29	36.89	37.57

For the proposed research work, classification is done using deep learning –DenseNet 169 classifier. Based on the result, this work gives benign and malignant characteristics corresponding to moderate lesions and highly suspicious lesions, depending on semantic values. These metrics are extracted automatically by the deep learning classifier during the process of feature extraction with respect to mean, contrast, correlation, energy value, etc. as shown in Table 2.

Table 2

Feature values extracted during classification

Image no	ABCD				GLCM				FOS
	Asymmetry	Border Irregularity	Colour	Diameter	Correlation	Energy	Contrast	Homogeneity	Mean	Standard deviation	Kurtosis	Entropy	Skewness
1	0.147	4.1437	Black	4.36	0.94750	0.8136	0.0086	0.99743	0.008	0.0106	6.7625	2.3845	2.0178
15	0.532	3.1479	Brown	4.68	0.98246	0.7468	0.00568	0.99654	0.017	0.0198	3.3475	2.7465	1.3498
142	0.149	5.1049	Red patch	2.64	0.94253	0.7748	0.00547	0.99674	0.019	0.0148	4.7156	2.7468	1.5468
154	0.167	3.6941	Red patch	1.98	0.97424	0.6432	0.00667	0.99576	0.016	0.0119	3.3380	2.9456	1.2465
277	0.119	3.5634	Brown	3.46	0.96347	0.7954	0.00845	0.99773	0.008	0.0107	7.1286	2.0564	2.2478
293	0.106	6.4587	Brown	4.86	0.96554	0.8476	0.00458	0.99431	0.010	0.0124	4.1874	2.5745	1.6054
365	0.149	3.1759	Black	2.22	0.98746	0.8214	0.00914	0.99337	0.006	0.0146	2.4956	2.4589	1.4789
390	0.348	4.9657	Black	1.08	0.94756	0.5744	0.00786	0.99646	0.007	0.0081	2.2846	2.7455	0.8546
418	0.453	4.4587	Pale red	5.18	0.91475	0.5596	0.00945	0.99775	0.011	0.0191	2.2569	2.2365	0.9475
266	0.210	4.6584	Brown	4.88	0.98472	0.5321	0.00514	0.99733	0.010	0.0127	4.164	2.8144	0.4297

Figure 2 shows the output of the input image, preprocessing, and segmentation stages for various images. The proposed OTFRS segmentation helps in discriminating between overlapping and non-overlapping dermoscopy lesion regions. Various training and testing lesion images are used by the classifier. Here, 10 lesion images are randomly chosen for the feature extraction process. The Recursive Pixel Histogram Equalization outperforms other existing methods with reduced error values.

Fig. 2

Detection of skin tumor region with Row 1: Input sample, Row 2: Preprocessing stage and Row 3: Segmentation stage.

Different segmentation methods with performance metrics: dice coefficient, sensitivity, jaccard index, and accuracy are compared in Table 3. Comparisons show that the proposed segmentation outperformed all other methods with an accuracy: 98.15%, a dice coefficient of 93.44%, a sensitivity of 98.77%, and a jaccard index of 88.93%. The proposed workflow results are shown in Figs. 3 5, which give the lesion OTFRS segmentation and DenseNet-169 classification for the benign and malignant categories from the entire dataset.

Table 3

Comparison of various segmentation methods

Methods	Dice		Jaccard	Accuracy
	Coefficient	Sensitivity	Index
MSCA [29]	0.8584	0.5643	0.8869	0.9365
FCN [30]	0.8637	0.5883	0.8295	0.9707
MFCN [31]	0.8524	0.6750	0.8349	0.9655
SSLC [32]	0.9137	0.7983	0.8681	0.9813
J-FCN [25]	0.9113	0.8961	0.8853	0.9534
PPBRS [27]	0.9263	0.9383	0.8949	0.9774
OTCRBS [27]	0.9341	0.9578	0.8849	0.9678
Our Method	0.9344	0.9877	0.8893	0.9815

Fig. 3

Classification of benign skin tumor region with Row 1: Input sample, BAD filtering, HSV colour conversion, Row 2: Histogram equalization, CSS, Proposed OTFRS segmentation and Row 3: Benign -warning sign.

Fig. 5

Classification of malignant skin tumor region with Row 1: Input sample, BAD filtering, HSV colour conversion, Row 2: Histogram equalization, CSS, Proposed OTFRS segmentation and Row 3: Malignant -warning sign.

This enriched process can even be applied to other tasks using the DLC –DenseNet 169 classifier based on extracted or newly pre-trained feature sets. The over-fitting of the classifier are prevented by two main criteria: (i) through proper data augmentation creating diverse training images through random transformations on existing images creating new set of images (ii) By choosing only the best features for training the classifier where the GLCM, FOS and ABCD values corresponding to correlation, energy, contrast and homogeneity, mean, standard deviation, kurtosis, entropy and skewness, asymmetry, border, color and diameter.

The smaller size, which can be less than or equal to 3 cm, is considered benign, and a size greater than the value is considered malignant. Benign tumor waning indicates low risk values. The malignant tumor warning shows high-risk lesion ranges. Table 4 gives a comparison of the proposed classifier with other methods based on error and accuracy values. Figure 4 gives the graph representation, which proves the improvement in research interpretation and research analysis.

Table 4

Classifier comparison - accuracy with error values

Classifier	BRT [25]	KNN [8]	SVM [7]	ANN [4]	SVM with BoVW [28]	FFNN [16]	Proposed DLC- DenseNet 169
Accuracy (%)	91.7	93	92.54	93.71	95.5	96.3	98.86
Error (%)	8.3	7	7.46	6.29	4.5	3.7	1.14

Fig. 4

Accuracy and error values - comparison.

Table 5 and Fig. 6 give sensitivity, specificity, precision, accuracy, and time value comparisons, which prove the effectiveness of the proposed research work. The sensitivity of the proposed classifier shows a value of 96.11%. The values of sensitivity should always be kept at their maximum values to maintain a steady state. However, this slight decrease in sensitivity compared to other techniques does not affect the accuracy, which is about 98.86%.

Table 5

Comparison: Performance values using different classifiers

Performance measure	ANN [4]	ResNet -50 [34]	BRT [25]	KNN [8]	SVM [10]	SVM-BoVW [28]	FFNN [16]	AlexNet [33]	GRU [35]	Proposed Classifier
Sensitivity (%)	92.34	–	88.73	92.33	92.75	96.48	97.21	95.60	96.81	96.11
Precision (%)	94.73	86	85.14	86.74	91.38	96.48	87.78	95.06	96.80	96.88
Specificity (%)	96.4	–	90.24	93.78	89.94	95.01	93.46	93.27	94.5	95.43
Accuracy (%)	93.71	84.87	91.7	93	93.11	95.5	96.3	98.70	97.1	98.86
Time (second)	46	–	39	43	29	34	24	–	–	21

Fig. 6

Comparison - classifier performance metrics.

Compared to other state-of-the-art methods, the proposed classifier’s accuracy is high, as shown in Table 5. 100% benign and 98.6% malignant classification is done successfully by the proposed DLC- DenseNet 169 classifier in Fig. 7.

Fig. 7

Deep learning Classification - Confusion Matrix.

6 Classifier- comparison

Effective comparisons of the skin lesion dataset are made with other classification systems. SVM-BoVW provides classification for the image categories. ANN, FFNN, and KNN perform classification based on two dissimilar classes. SVM and BRT classify two classes without any image features and without any of the aforementioned data on image features. The greater the error range, the lower the accuracy. A comparison of classifiers for TP, TN, FP, and FN is shown in Table 6, and it is proven that the proposed classification performs better on skin lesion classification with a high true positive value.

Table 6
Comparison of classifier performance - confusion matrix

Classification TP[% ] FP[% ] FN[% ] TN[% ]

ANN [4] 41.2 52.8 47.2 58.8

BRT [25] 44.7 55.7 44.3 55.3

KNN [8] 45.8 55.9 44.1 54.2

SVM [10] 42.9 59.1 40.9 57.1

SVM-BoVW [28] 46.7 63.7 36.3 53.3

FFNN [16] 43.2 52.5 47.5 56.8

Proposed 48.7 65.3 34.7 51.3

Classification	TP[% ]	FP[% ]	FN[% ]	TN[% ]
ANN [4]	41.2	52.8	47.2	58.8
BRT [25]	44.7	55.7	44.3	55.3
KNN [8]	45.8	55.9	44.1	54.2
SVM [10]	42.9	59.1	40.9	57.1
SVM-BoVW [28]	46.7	63.7	36.3	53.3
FFNN [16]	43.2	52.5	47.5	56.8
Proposed	48.7	65.3	34.7	51.3

These comparative results demonstrate that SVM-BoVW and FFNN have 96.2% and 97.1% accuracy, respectively. KNN has 93.54%, BRT has 92.7%, and ANN has 93.21% accuracy. However, the proposed DLC-DenseNet 169 classification technique gives 98.86% accuracy. It holds 5.6 s per slice of computation time. This novel research classification helps dermatologists detect skin lesions. The loss function curves versus the iteration values on testing and training sequence is given in Fig. 8.

Fig. 8

Loss function curve versus iteration with DenseNet -169 classifier.

Three sets of skin lesion images are trained to minimize the loss functions. With the training of these samples the loss functional values maintains to be minimum.

High TP values with higher accuracy are obtained with the proposed DLC methods, as shown in Fig. 9. This curve with a high true positive shows increased accuracy with low error values when compared to other classifiers in various fields. Table 7 shows the DenseNet-169 performance comparison based on the testing and training data with 70-30%, 80-20%, and 90-10% for two different classes of benign and malignant skin tumors, respectively.

Fig. 9

Comparison ROC plot of classifier.

Table 7

Deep learning classification -DenseNet-169

Classifier	Data (Testing- Training)	Class	Sensitivity	Specificity	Precision	Accuracy
DenseNet –169	70% -30%	Benign	0.956	0.964	0.924	0.982
		Malignant	0.965	0.912	0.957
		Avg. Value	0.960	0.938	0.941
	80% -20%	Benign	0.929	0.957	0.951	0.984
		Malignant	0.963	0.977	0.947
		Avg. Value	0.946	0.967	0.949
	90-10%	Benign	0.973	0.974	0.984	0.982
		Malignant	0.941	0.966	0.954
		Avg. Value	0.957	0.97	0.969

Table 8

Nomenclature

ABCD	Asymmetry, Border, color, Diameter
AI-DLC	Artificial Intelligence - Deep Learning Classification
AMBE	Absolute Mean Brightness Error
AND	Anisotropic Notch Diffusion
ANN	Artificial Neural Network
AV	Asymmetric Values
BADF	Boosted Anisotropic Diffusion filter
BRT	Bayesian regression Tree
CPU	Central Processing Unit
CSS	Contour superpixel segmentation
CT	Computed tomography
DICOM	Digital Imaging and Communications in Medicine
FCN	Fully convolutional Network
FFNN	Feed Forward Neural Network
FOS	First Order Statistics
GHz	Giga Hertz
GLCM	Gray level Co-occurrence Matrix
GRU	Gated Recurrent Unit
HSV	Hue Saturation Value
ICC	International Cancer Center
ISIC	International Standard Industrial Classification
JPEG	Joint Photographic experts Group
KNN	K- Nearest Neighbor
MBE	Mean Brightness Error
MF	Main Features
MFCN	Multi-scale Fully convolutional Network
MSCA	Multi-scale contextual attention network
MSE	Mean Square Error
OTCRBS	Online Tiger Claw Region Based Segmentation
OTFRS	Online Tigerclaw Fuzzy Region Segmentation
PC	Personal Computer
PNG	Portable Network Graphics
PPBRS	Prioritized patch Based region Segmentation
PSNR	Peak Signal to Noise Ratio
RGB	Red Green Blue
ROI	Region of Interest
RP	Recursive Pixel
RPHE	Recursive Pixel Histogram Equalization
SF	Structural Features
SSLC	sequential search based on linear collisions
SVM	Support Vector Machine
SVM-BoVW	Support Vector Machine –Bag of Visual Words
U-Net	U- shaped Network architecture

7 Conclusion

In this paper, an automatic system of computer based classification along with segmentation is implemented to assist radiologists with skin lesion dermoscopic images. Our research work is based on deep learning classification for benign and malignant lesion images of tumor cells. We propose to prove that DLC with DenseNet 169 has high accuracy with OTFRS segmentation under abnormal and normal conditions. The proposed Online Tigerclaw Fuzzy Region Segmentation provides effective detection and segmentation of lesion images. 5.6 s per slice of computation time is obtained using MATLAB software on a laptop with an Intel Core i7. This novel research classification helps dermatologists detect skin lesions. Thus, the proposed OTFRS method has accuracy for benign: 98.9 % and malignant: 97.4%. Deep Learning Classification-DenseNet 169 is 100% and 98.86% for benign and malignant, respectively. While achieving the improvement in time for the proposed method, the sensitivity decreases gradually. In spite of a slight decrease in the value of sensitivity, an improvement in accuracy is still achieved. The proposed research work is effective in the prior detection of skin lesion cells when compared to other state-of-the- art methods.

Footnotes

Acknowledgments

The authors acknowledge their role in the creation of the free publicly available ISIC skin tumor database and the International Cancer Centre (ICC), Neyyoor, Tamil Nadu for providing the skin tumor database used in this research work.

References

Kani

M.A.J.M.

, Parvathy

M.S.

, Banu

S.M.

and Kareem

M.S.A.

, Classification of skin lesion images using modified InceptionV3model with transfer learning and augmentation techniques, Journal of Intelligent & Fuzzy Systems Preprint: 1-15.

Ding

and Wang

, Efficient Unet with depth-aware gated fusion for automatic skin lesion segmentation, Journal of Intelligent & Fuzzy Systems 40(5) (2021), 9963–9975.

Harangi

, Skin lesion classification with ensembles of deep convolutional neural networks, Journal of Biomedical Informatics 86 (2018), 25–32.

Ercal

, Chawla

, Lee

H.C.

and Moss

R.H.

, Neural network diagnosis of malignant melanoma from color images, IEEE Transactions on Biomedical Engineering 41(9) (1994), 837–845.

Thomas

S.M.

, Lefevre

J.G.

, Baxter

and Hamilton

N.A.

, Interpretable deep learning systems for multi-class segmentation and classification of non-melanoma skin cancer, Medical Image Analysis 68(101915) (2021).

Masood

, Al-Jumaily

A.A.

and Adnan

, Development of automated diagnostic system for skin cancer: Performance analysis of neural network learning algorithms for classification,Springer, Artificial Neural Networks and Machine Learning -. Computer Science 8681 (2014).

Harika and Lakshmi

, Segmentation and classification of skinlesions from dermoscopic images, Journal of Scientific andIndustrial Research (JSIR) 80(04) (2021), 328–335.

Waibel

D.J.E.

, Boushehri

S.S.

and Marr

, Instant DL: An easy-to-use deep learning pipeline for image segmentation and classification, BMC bioinformatics 22(1) (2021), 1–15.

Afshar

, Singla

and Khandelwal , Automatic detection of pectoralmuscle using average gradient and shape based feature, Journalof Digital Imaging 25(3) (2012), 387–399.

10.

Satheesha

and Dhruve

, Melanoma is skin deep: A 3D reconstruction technique for computerized dermoscopic skin lesion classification, IEEE Journal of Translational Engineering in Health and medicine 5 (2017), 1–17.

11.

Khan

M.A.

, Sharif

, Akram

, Damasevičius

and Maskeliūnas

, Skin lesion segmentation and multiclass classification using deep learning features and improved moth flame optimization, Diagnostics 11(5) (2021), 811.

12.

Seiffert

, Khoshgoftaar

T.M.

, Hulse

J.V.

and Napolitano

, Rusboost: A hybrid approach to alleviating class imbalance, Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions 40(1) (2010), 185–197. doi:10.1109/TSMCA.2009.2029559.

13.

Vese

L.A.

and Chan

T.F.

, A multiphase level set framework for image segmentation using the Mumford and Shah model, International Journal of Computer Vision 50(3) (2002), 271–293.

14.

Singh and Mary , The detection of pulmonary nodules in CT images using heuristic approach segmentation and classification, ARPN J Eng Appl Sci 11(9) (2016).

15.

Liu

, Ouellette

, Jamgochian

, Liu

and Rao

, One-class machine learning classification of skin tissue based on manually scanned optical coherence tomography imaging, Scientific Reports 13(1) (2023), 867.

16.

Manimurugan

, Hybrid high performance intelligent computing approach of CACNN and RNN for skin cancer image grading, Soft Computing 27(1) (2023), 579–589.

17.

Alenezi

, Armghan

and Polat

, A novel multi-task learning network based on melanoma segmentation and classification with skin lesion images, Diagnostics 13(2) (2023), 262.

18.

Alheejawi

, Xu

, Berendt

, Jha

and Mandal

, Novel lymph node segmentation and proliferation index measurement for skin melanoma biopsy images, Computerized Medical Imaging and Graphics 73 (2019), 19–29.

19.

Lather

and Singh

, Tumor segmentation from brain MR images using STSA based modified K-means clustering approach, Journal of Intelligent & Fuzzy Systems Preprint (2022), 1–17.

20.

Saba

M.A.K.

, Tanzila

A.R.

and Larabi

, Region extraction and classification of skin cancer: A heterogeneous framework of deep CNN features fusion and reduction, Journal of Medical Systems 43(9) (2019), 289.

21.

Sreelatha

and Prasad

, Early detection of skin cancer using melanoma segmentation technique, Journal of Medical Systems 43(7) (2019), 190.

22.

Mangalakshmi

, Devi

and Syamala , Lung nodule classification using deep features in CT images, Int J Adv Res Comput Sci Softw Eng 6(3) (2016).

23.

Wang

, Yan

, Tang

, Yang

and Chen

, Multiscale feature fusion for skin lesion classification, BioMed Research International 2023 (2023).

24.

Jianbo

, Local and nonlocal preserving projection for bearing defect classification and performance assessment, IEEE Trans Ind Electron 59(5) (2012), 2363–2376.

25.

Karthik

B.U.

and Muthupandi

, SVM and CNN based skin tumor classification using WLS smoothing filter, Optik 272 (2023), 170337. https://isic-archive.com.

26.

Ashwini

and Murugan

, Automatic skin tumor segmentation using prioritized patch based region- a novel comparative technique, IETE Journal of Research 17(1) (2020), 177.

27.

, Niu

, Sisternes and Rubin , Bag-of-visual-words models for adult image classification and filtering, In Pattern Recognition, 19th International Conference on ICPR, 2008, pp. 1–4.

28.

Chauhan

and Goyal

, A Multi-path CNN for automated skin lesion segmentation, In 2021 International Joint Conference on Neural Networks (IJCNN), IEEE, 2021, pp. 1–9.

29.

Kaymak

, Kaymak

and Ucar

, Skin lesion segmentation using fully convolutional networks: A comparative experimental study, Expert Systems with Applications (2020), 113742.

30.

Hasan

Md.K.

, Dahal

, Samarakoon

P.N.

, Tushar

F.I.

and Martí

, DSNet: Automatic dermoscopic skin lesion segmentation, Computers in Biology and Medicine 120 (2020), 103738.

31.

Khouloud

, Ahlem

, Fadel

and Amel

, W-net and inception residual network for skin lesion segmentation and classification, Applied Intelligence (2021), 1–19.

32.

Ashwini

and Kavitha

, Automatic skin tumor detection 664 using online tiger claw region based segmentation–A novel 665 comparative technique, IETE Journal of Research (2021), 69(6).

33.

Hosny

K.M.

, Kassem

M.A.

and Fouad

M.M.

, Classification of skin lesions into seven classes using transfer learning with AlexNet, Journal of Digital Imaging 33 (2020), 1325–1334.

34.

Mehra

, Bhati

, Kumar

and Malhotra

, Skin cancer classification through transfer learning using ResNet-50, In Emerging Technologies in Data Mining and Information Security: Proceedings of IEMIS,Volume 2, 2020, pp. 55–62.

35.

Ahmad

, Usama

, Ahmad

, Khatoon

and Alam

C.M.

, An ensemble model of convolution and recurrent neural network for skin disease classification, International Journal of Imaging Systems and Technology 32(1) (2022), 218–229.